A system for automatic, reliable, semi-instantaneous estimation of avalanche activity is presented in an attempt to check deterministic models and improve the surveillance of risk zones. The principle is the seismic detection of avalanches. Two experimental sites equipped with standard seismological equipment are described. Avalanche seismic signals are recorded, as well as many extraneous signals of natural or human origin. After several years of a posteriori identification of the signals, we are now able to set up criteria for the automatic recognition of non-avalanche signals. We have therefore developed an automatic analysis/decision system to discriminate between avalanche signals and others. This system worked satisfactorily in pre-operational conditions during the winter of 1995-96. The results are presented and compared with other data related to avalanche activity. Although there is still room for improvement, our system seems to be able to estimate avalanche activity better than human visual observations.