LOCAL-DENSITY EVALUATION AND TRACKING OF MULTIPLE OBJECTS FROM COMPLEX IMAGE SEQUENCES

In the present paper, multiple-hypothesis static knowledge modelling is introduced in the distributed inference mechanism of Extended Kalman Filters presented in [l]. The extension from single to multiple modelling based on spatial constraints the system performances to be improved and the system functionalities to be extended (in terms of types and number of parameters to be estimated fiom acquired data): the present system provides in real-time not only crowding density estimation but also people tracking in real-life complex environments. The target has a major role in the field of surveillance of complex environments. It is reached by extracting fom sequences of images a set of significant features, correlated with the number of people present in the monitored scene. The proposed new approach consists in providing each feature-extractor node with several possible non-linear models and with a classincation procedure for selecting the most suitable one. A probabilistic approach has been developed in order to track the monitored groups of people by means of a graph-based time and space knowledge representation