Braking Detection and Prediction with Inter-vehicle Distance Estimated from Driving Videos

Heavy braking is a common driving behaviour that indicates a possible traffic hazard. To avoid traffic accidents, detecting and predicting braking events are essential to inform the driver to take action beforehand. Recently, machine learning driven by probe data has been applied to analyze braking events. However, the causes of heavy braking are complex, and utilizing solely the probe data from drive recorders is insufficient for effectively analyzing the braking events. In this study, we first present a comprehensive analysis of braking causes through human annotation from driving videos. Through manual annotation, we distinguish the misjudgments of braking events introduced by the accelerometers. We subsequently build models to detect and predict braking events based on the annotation. To increase the performance, we propose to utilize instance segmentation and monocular depth estimation to approximate the inter-vehicle distance from driving videos, and treat it as a feature in the machine learning model in addition to probe data. Experimental results demonstrate that maintaining inter-vehicle distance is an important cause of braking and our models improve the performance compared to previous probe-data-only models by incorporating the distance information.