Automobile safety can be improved by detecting crash quickly, and thereby providing additional time to deploy new safety devices such as slower airbag inflation rates, alarms, seat belt pre-tensioners, window closing, seat stiffeners etc. This requires a discriminative, fast and robust crash sensing system that evaluates crash data and discriminates crash types over a wide range of scenarios. Prompt crash detection not only decreases the mortalities and injuries, but also reduces unnecessary airbag deployments. Studies have shown that unnecessary airbag deployments can cause greater injuries than minor crashes would causes.

The state of the art Smart Airbags faces two challenges, i.e., quick crash sensing times and discrimination among crashes over a wide range of conditions. Current crash sensing algorithms utilize deceleration or speed change to estimate the severity of crash. Since neither deceleration nor speed change is linear, the resulting algorithms cannot differentiate crash situations, such as low speed rigid barrier versus high-speed deformable barrier, in a timely and robust manner. Moreover, the conventional algorithms are developed using ad hoc processes. The resulting algorithm is not general. We present a Single Point Crash Sensing System based on machine learning principles. Compare to conventional system, our system has the following advantages: (1) Centralized. The Single Point Crash Sensing architecture reduces the cost of components and wiring, and also decreases the vulnerability of the system; (2) Fast. Current frontal impact sensing systems generally detect a crash condition within 15-25 milliseconds; and side impact sensing systems generally discriminate side impacts within 6-13 milliseconds. Our system can detect all kinds of crashes within 6 milliseconds; (3) General. The modularized algorithm present in this thesis is a general approach, which can be easily adapted to different vehicle models. (4) Discriminative. The experiment results indicate that the system maintains high classification accuracy in data streams, hence providing an efficient solution to the classification task. (5) Reliable. The experiment shows that our systems are tolerant of noise. Both HMM Crash Detector and SVM Crash Detector can distinguish crashes from non-crashes in 6-7 ms even if the injected noises reach 100% level.