A coarse-to-fine vehicle detector running in real-time

Abstract

An algorithm, that allows the detection of vehicles through the analysis of images, has nowadays a number of applications for safety, increasing the drivers awareness in potential dangerous situations. The development of a reliable and robust vehicle detection is an open challenge for the computer vision society that continues to be largely investigated. This work describes a complete framework to detect front vehicles and estimate their position, speed and their lane-level localization, with a focus on performance and real-time capabilities. The described system exploits the combination of an AdaBoost Soft-Cascade with Integral Channel Features, an efficient and precise multiscale approach, and a final stage of matching and tracking of detected vehicles. The full system has been evaluated on a public dataset, and it is being tested in real situations, nowadays exhibiting performances compatible with the use in practical applications.