Multisignal neural classification and temporal distortions: application to the detection of driver vigilance.

Summary The automatic detection of driver's vigilance decreases is considered from the signals describing the vehicle driving. As most of the phenomena that reflect a human intervention, temporal distortions distort these signals and mask the information sought. A first approach consists in temporally shifting the signals. A shift energy quantifies the effort required to realign two curves. The minimization of this energy is achieved by dynamic programming. A second approach consists in using a coding of the signals adapted to their nature. Based on a wavelet transform, this representation quantifies the size distribution of the oscillations at each scale. This coding is robust to temporal distortions and achieves a significant compression of the data. A family of primary vigilance classifiers is built: each classifier bases its decisions on the coding of a given signal at a given scale. The classification is established by an artificial neural network. The weights of the network are selected after scanning several architectures. A multi-sensor and multi-scale hypovigilance detection is finally built by optimal combination of the primary classifiers.