Direct Automatic Detecting Technique of Traffic Accident Based on Acoustic Signals

Non-recurrent traffic congestion is mainly the result of traffic incidents such as accidents, vehicle breakdowns or any other incident. This paper focuses on traffic accident as it is the dominant cause. Faster detection of traffic accidents is of extreme importance for fast rescuing the wounded and quickly clearing the bottlenecks. The automated accident detection algorithm can largely be categorized into two groups. The algorithms in the first group are based on the symptoms of the accidents, such as changes in traffic volume, speed, and occupancy, and those in the second group are directly based on accident features such as sound and image. The former is imprecise and not Real-time because the arithmetic depends on road circumstance and historical data. Image-processing based accident detection is possible, but it is neither cost effective nor computationally efficient compared with audio-based detection methods. This paper studied direct automatic detecting technique of traffic accident based on processing and recognizing acoustic signals. A database consisting of sounds from vehicle crashes, thunder, and sounds of running normally vehicle was created. We abstract the features by wavelet package and compute the wavelet coefficients as a feature vector for input to the classification system. A BP neural network is used to classify these features into categories of crash and non-crash events. The classification testing results achieved 95% accuracy. These results show that the feature of radiated noise could reflect vehicle running state and could be an effective measurement for traffic accident directly detecting.