Abstract<br />The main subject of this thesis is motion prediction. The problem is approached from the hypothesis that the dynamic and kinematic properties of objects such as pedestrian and vehicles do not suffice to predict their motion in the long term. Instead, the work presented here, in scribes itself in a new family of approaches which assume that, in a given environment, objects do not move at random, but engage in “typical motion patterns”, which may be learned and then used to predict motion on the basis of sensor data. In this context, this thesis focuses in three fundamental questions: modeling, learning and prediction.<br /><br />Modeling. <br /><br />This thesis is based on Hidden Markov Models, a probabilistic framework, which is used as a discrete approximation to represent the continuous state-space in which motion takes place. The main originality of the approach lies in modeling explicitly the intentions which are at the origin of “typical motion patterns”. This is achieved through the using of an extended space, which adds the state that the object intends to reach to the other “classic” state variables, such as position or velocity.<br /><br />Learning. <br /><br />The main problem of existing approaches lies in the separation of model learning and utilization in two distinct stages: in a first phase, the model is learned from data; then, it is used to predict. This principle is difficult to apply to real situations, because it requires at least one example of every possible typical pattern to be available during the learning phase. To address this problem, this thesis proposes a novel extension to Hidden Markov Models which allows simultaneous learning and utilization of the model. This extension incrementally builds a topological map – representing the model's structure – and reestimates the model's parameters. The approach is intended to be general, and it could be used in other application<br />domains such as gesture recognition or automatic landmark extraction.<br /><br />Prediction. <br /><br />In this context, prediction is carried on by using exact Bayesian inference algorithms, which are able to work in real time thanks to the properties of the structure which has been learned. In particular, the time complexity of inference is reduced from O(N 2 ) to O(N) with respect to the number of discrete states in the system. <br /><br />All of the results obtained on this thesis have been implemented and validated with experiments, using both real and simulated data. Real data has been obtained on two different visual tracking systems: one installed over a parking lot, and the other installed at INRIA's entry hall. For synthetic data, a simulator has been developed in order to facilitate the conduction of controlled tests and the study of larger environments than for real data.