Driving Event Detection by Low-Complexity Analysis of Video-Encoding Features

Abstract

All standard video-encoding algorithms rely on differential encoding with motion compensation to improve the compression. When a video from a front-facing camera onboard a vehicle is compressed, the information computed for compression purposes, in particular motion vectors, can be effectively used to gain some understanding of the driving dynamics and eventually to support driver decisions and improve driving safety. In this chapter an algorithm that can use such side information to detect a number of driving events is presented. Numerous potential applications are envisaged. Since video-encoding software and hardware are usually strongly optimized, it is possible to implement the proposed algorithms in battery-powered embedded devices with strict limits on processing capabilities such as camera-equipped mobile phones mounted on the car dashboard and consequently allow different types of low cost vehicles, which in most cases do not include cameras as a standard equipment, to be fitted with at least a warning device with very low cost. If the video is captured in the context of a video surveillance scenario, differentiating the events could be used to automatically decide which portion of the video should be transferred to a remote monitoring center thus optimizing network resources usage and costs.