The objective of the research is to realize an Intelligent Vision Machine, developed in hierarchical levels, for the detection of illegal activities of stealing in archaeological sites. The first level detects the activities, identified as significant movements, in the area of interest giving rise to an alarm at a remote control station where human operators take decisions about the event by using also the transmitted images. The second level analyses the observed activities in order to reduce the number of false alarms sent to the operators. The filtering of alarms works at different levels with increasing complexity. The low level detects the moving objects. Some classes of objects such as people, animals, cars, etc., are analysed more in detail in order to avoid alarms due to not illegal actions such as animals moving in the area of interest. The class of people is deeply analysed to distinguish the ordinary actions (transit, agricultural works, ...) from the illegal ones (survey of the terrain, excavation, stealing, ...). This analysis avoids to send alarms for clearly harmless actions. The third level, which characterizes the human activities, consists of two steps: the detection of the posture in each image and the temporal analysis of postures in the whole sequence of images. The first step detects the attitude of the body of the detected person and classifies it into three fundamental postures: standing, squatted and bent. The algorithms are based on the horizontal and vertical projections of the binary shape of the person and on unsupervised clustering able to set the parameters during the observation of the image sequences without the human intervention. The second step analyses statistically the temporal sequence of the postures and distinguishes between legal and illegal actions. The Hidden Markov Models are the theoretical mathematical-statistical model used for the implementation of the vision machine. They allow to model first the hostile behaviors by analysing the postures (training phase) in sample sequences and second to automatically recognize the activities analysing the acquired sequences. The sensors used for the observations are specific visual sensors able to work under different lighting conditions (daylight, night) and different meteorological conditions (visual, infrared and intensifying cameras). The visual sensors are organized to face two contemporary functions: global analysis of the area of interest detecting and tracking the movements and detailed analysis of the single sources of movement.
Immagini:
