Crash Detection System Using Hidden Markov Models

This paper presents the design of a crash detection system based on the principles of continuous-mode Hidden Markov Models (HMM) with real-valued emission parameters. Our design utilizes log-likelihood for optimizing HMM parameters including the number of states in the model and the accelerometer crash-pulse buffer size resulting in lower costs and complexity of the crash detection system. Cross validation technique based on Jackknifing is utilized to estimate the crash pulse detection rate for a variety of crash events involving rigid as well as offset deformable barriers with head-on and oblique angle impacts. The system is simulated using Matlab and Simulink, and the proposed model is able to accurately classify crash-events within 10 ms from the time of the impact.