Performance boosting of a pedestrian detector using multiple cameras

Pedestrian detection is a key problem in computer vision; for autonomous vehicles a pedestrian detection system has to operate in real time and to be highly accurate. In the past twenty years there was a magnificent leap in the detection accuracy of objects and pedestrians in single frame images. However, the results are far from perfect, especially for the fast pedestrian detectors which have relatively low detection accuracy. This paper introduces an approach for improving the results achieved today by a single camera pedestrian detection system. By applying the pedestrian detector on multiple cameras, it is possible to find a correlation between the detections in the two cameras. This allows to discard false positives and to estimate the distance using triangulation for 3D estimation and tracking. The results received show considerable improved performance in the detection rate as well for the detection precision.