Toshihiko Itoh

Prediction of Driving Actions from Driving Signals

A spoken dialogue system for car-navigation systems may be able to provide more natural and smoother communications but it must also cause safety problems. One of these problems is distraction whereby machine operation and voice conversations influence the driver. Even the use of a simple speech interface may affect the driving operation. We consider that a spoken dialogue system which can understand the drivers situation and change its dialogue rhythm according to that situation would be safe as part of a car-navigation system. For this to be possible, the system needs to predict and recognize drivers actions from environmental information such as driving signals. In this chapter, we report the results of an experiment on predicting driver actions. The action prediction system uses HMM-based pattern recognition only on driving signals and does not use position information. Its best driving action prediction accuracy was 0.632.

A point-pass-based action prediction method

The paper describes a basic idea on how to realize an intelligent learning room system. Such a system needs to have a dynamic adaptive capability for each user. We have proposed a method to predict user action using inductive learning with N-gram. The system based on our proposed method is able to acquire rules automatically from data pairs through inductive learning. As unified with N-gram, the system demonstrates a high predictive accuracy. However, the acquired rules express the users habits and preferences. Consequently, it is possible that the system adapts dynamically to each user. The user needs to proof-read the errors in the prediction results. Therefore, the prediction ability improves. As a result, the number of errors decreases. This paper unifies N-gram and inductive learning to develop the point-pass-based prediction system. The system was found to have good accuracy of which the highest prediction accuracy was about 89.3%. The system was proved to have high dynamic adaptive ability.

Evaluation of action prediction method using inductive learning with N-gram

Being society aging, an intelligent room is needed for the aged or handicapped. The important ingredient of such a system is how to predict the next action. In this paper we describe how to solve the problem of predicting inhabitant action in an intelligent room that we called learning room. We have proposed a method to predict user action using inductive learning (IL) with N-gram. The system based on our proposed method is able to acquire the immanent causality rules automatically from data pairs by means of IL. Since our system unified IL and N-gram, it demonstrates good accuracy for the simulated data. The system showed high dynamic adaptive capability.