Title: Motion learning and prediction for vehicles and pedestrians

Authors: Dizan Vasquez, Thierry Fraichard, Christian Laugier

Speaker: Christian Laugier

Affiliation: INRIA, Rhone Alps, France  (Team eMotion/GRAVIR/CNRS)

Abstract:
Until recently, most motion prediction techniques have been based on kinematic or dynamic models that describe how the state of an object evolves over time when it is subject to a given control . Although these techniques are able to produce very good short-term predictions, their performance degrades quickly as they try to see further away in the future. This is especially true for humans, vehicles, robots, animals and the like, which are able to modify their trajectory according to factors (eg perception, internal state, intentions, etc.) which are not described by their kinematic or dynamic properties.

To address this issue, a different family of approaches has emerged recently. It is based on the idea that, for a given area, moving objects tend to follow typical motion patterns that depend on the objects’ nature and the structure of the environment. Such approaches operate in two stages:

1. Learning stage: observe the moving objects in the workspace in order to determine the typical motion patterns.
2. Prediction stage: use the learnt typical motion patterns to predict the future motion of a given object.

But existing learning techniques have a drawback: they use a “learn then predict” approach, meaning that the system goes through a learning stage where it is presented with a set of observations (an example data set) from which it builds its pattern models. Then, the models are “frozen” and the system goes into the prediction stage. The problem with this approach is that it makes the implicit assumption that all possible motion patterns are included in the example dataset, which, is a difficult condition to meet. In contrast, we are exploring the “learn and predict” approach. That is, learning is an incremental process, which continuously refines knowledge on the basis of new observations which are also used for prediction.

We start from the hypothesis that objects move in order to reach specific places of the environment (ie goals). Hence, motion patterns are represented using a Hidden Markov Models (HMM) and integrating the goal in the state variable. Thanks to this, it is possible to predict the object’s destination as a part of the state estimation (ie filtering). The state also may be predicted at any given time in the future using probabilistic inference. Our main contribution is a novel approach to learning the structure and parameters of probabilistic models using a self organizing network based on the Instantaneous Topological Map (ITM) algorithm. The idea is that the structure of the HMM should reflect the structure of the environment. Since the structure is discovered in an incremental fashion, the HMM structure evolves as well, giving place to what we have called a Growing Hidden Markov Model (GHMM).

A major advantage of the use of the enhanced ITM algorithm is that, in general, the number of edges in the transition graph is linear with respect to the number of discrete states in the model, as opposed to quadratic for a full connected HMM. At the same time, the structure is much more expressive than the restrictive linear (order 2) topologies obtained by other methods. Yet, the structure is simple enough to enable the use of exact inference algorithms in real-time, thus constituting an effective “learn and predict” approach.