Radar-Vision Based Vehicle Recognition with Evolutionary Optimized and Boosted Features

Abstract

We present a real-time monocular vehicle detection and recognition system for driver assistance based on the fusion of data from a radar and a video sensor. The radar data is used both for narrowing down the size of the search area for vehicle rears on the video image and for the distance measurement of the vehicles in front. Using the passive video sensor a radar object is verified and the width as well as the lateral position of the vehicle are determined. The contribution of this work is threefold: At first, we present and apply a methodology for developing a novel evolutionary optimized symmetry measure. Secondly, we demonstrate a vehicle detection and recognition algorithm consisting of two steps: hypothesis generation using a detector based on a set of Haar-like filters and an AdaBoost learning algorithm and hypothesis verification using an evolutionary optimized and biologically motivated vehicle recognition system. Finally, the performance of both the individual components and the complete vehicle detection and recognition system is evaluated by not only using classical confusion matrices but also giving information on the accuracy of the width and lateral position sensing. Our experimental results demonstrate a robust and real-time system trained and tested on more than 30,000 images.