ABSTRACT
Artificial neural networks (ANNs) are among the soft computing tools most frequently adopted for several video surveillance tasks due to their well-known advantages, such as adaptivity, parallelism, and learning [43]. Indeed, an ANN can modify its connection weights using some training algorithms or learning rules; by updating the weights, the ANN can optimize its connections to adapt to changes in the environment. Moreover, when the information is fed to an ANN, it is distributed to different neurons for processing, and neurons can work in parallel and synergistically, if they are activated by the inputs. The capability of neural networks in emulating many unknown functional links
for Video
by learning offline a limited set of representative examples allows one to infer a function from observations. This allows one to learn representations of the input that capture the salient input distribution features. In the general context of data classification, the neural network approach is particularly attractive, as it overcomes the difficulty of defining a single statistical model for different data types, which is the main problem associated with most conventional methods based on multivariate models. Also, neural approaches based on pyramidal structures have been pursued in the field of image processing and recognition, due to the compactness of multi-scale representations, which produce good textural features for landscape characterization, also providing support of efficient coarse-to-fine search [5]; efforts toward developing pyramid-based neural network techniques for recognition purposes arise as a way to handle problems of scaling with input dimensionality.
