Rash Driving Detection Using a Frontal View Camera in Cars

Many times accidents and unusual traffic congestion takes place due to careless and impatient driving. To increase awareness and promote safe driving, rash driving instances needs to be identified. To accomplish this, we propose an in-car visual-analysis based rash driving monitoring system which uses frontal-view video. This work combines two components which were trained independently. First component being vehicle motion detection (VMD), which takes frontal-view driving video as input. It predicts two most important vehicle motion parameters (forward and angular velocity). For this, we have adapted an existing architecture (proposed by Zhang et al.) and improved motion parameters prediction. A video dataset has been collected using android smartphone (Pilotguru) as capturing rash instances was necessary. Second component is rash driving instance classification model. For training this, Normal and Rash instances are identified from video and corresponding motion parameters (from video dataset ground truth) are combined in 6 second sequence creating a data-point. These data-points were annotated, resulting in rash dataset. For System validation and testing, given a frontal-view driving video, motion parameters were predicted using first component, which were then sequenced as 6 seconds clips and fed to second component predicting if the clip was rash or normal.