| Abstract [eng] |
Driving behavior is determined by individual driving capability features to control particular type of vehicle. These features form particular driving style, which depending on application may be categorized as: aggressive and safe driving, eco and fuel-demanding driving, sober and drowsy driving etc. Nowadays it is useful for transport companies to analyze driving behavior of their employees (or clients), because driving behavior affects vehicle life time, vehicle failure intensity, quality of transport services. Aggressive driving are considered all negative driving features that damage vehicle, increase risk of traffic accident, worsen transport services. In this work a detail driving behavior analysis is performed using inertial measurement sensor – 3- axis G-sensor (accelerometer) signal information, which is attached to vehicle. There are to objectives of this research: recognition of aggressive and safe driving behavior when signal information is obtained from vehicle that is driven in aggressive and safe driving styles; driving behavior identification (driver classification) when signals are obtained from public transport bus that is driven by two drivers (one at a time). Signal processing analysis is performed and algorithms formed that reduce noise level and detect outliers in signal by combining moving median and average filters. Also a special signal processing is done on public transport bus signals when stopping time must be extracted and deleted from raw signal. This is performed by applying Gaussian mixture models. Algorithms of feature extraction from accelerometer signal are formed using both time and frequency domain information. From time domain the following features are extracted: quantiles of signal; difference between quantiles; correlation between signals; data threshold violation intensity; standard deviation of jerk signal (first order differences of original signal). In the frequency domain relative power spectral density in particular frequency band are extracted as features for various frequency bands. Experimental analysis of aggressive and safe driving classification performance is investigated when four classification algorithms are used: support vector machines, neural networks, random forest and k-nearest neighbor algorithms. Feature selection using filter methods and feature transformation using principal components is perform in order to study whether this could improve classification accuracy. The results of aggressive and safe driving classification show that using 10 sorted features ~95% accuracy is achieved (that is greater than using all features) and using only 4 features extracted from X axis only ~91.5% accuracy is achieved. When performing public transport driver behavior classification the same all 78 features are extracted from public transport bus accelerometer signals and after feature extraction, a feature selection step is performed again together with driver classification. The results show that when using all 78 features, driving identification accuracy of ~79.4% is achieved, when using best 20 sorted features accuracy of ~79.2% is obtained and by using best 7 features, driver identification rate of ~77% is achieved. 6 of all 7 features are extracted from X axis signal alone. After aggressive and safe driving classification and public transport bus driving behavior classification it was noticed that both feature sets roughly match and are alike. So the conclusion is that differences of driving behavior are conditioned by different level of aggressiveness. |
