Intelligent Algorithms for Early Detection of Automotive Crashes

In this paper we describe the methodology for building crash pulses using accelerometer crash pulses. This procedure is applied to a set of 9 Chevrolet Cavalier crash pulses derived both from rigid and offset-deformable barrier collisions. Five (5) of these are used to train a 7-state, fully-connected HMM. The HMM model acquired from the training pulses is next utilized for classification of the remaining four crash pulses that have been inter-mixed with random pulses of comparable strength. The log-likelihood of deploy pulses was around -200, whereas the log-likelihood of non-deploy pulses ranged between -2000. As the public databases provide us crash pulses only for airbag deployment events, we have tested our algorithm using randomly generated pulses as negative control. The statistical characteristics of the randomly generated pulses are made to approximate that of the true pulse. The log likelihood of the crash pulses is very different from that of random pulses, which means we can distinguish a crash pulse from the random pulse only with the first 10 milliseconds' information. The classification of the crash and non-crash pulses was thus achieved without ambiguity.