Surround view pedestrian detection using heterogeneous classifier cascades

Abstract

Pedestrian detection is of particular interest to the automotive domain, where an accurate estimation of a pedestrian's position is the first step towards reliable collision avoidance systems. Driven by rapid advances in technology, several systems to detect pedestrians in front of a moving vehicle have been proposed in recent years. This paper introduces a novel pedestrian detection system for low-speed driving scenarios, capable of detecting pedestrians in a 360-degree fashion around the vehicle. Detected pedestrians are displayed to the driver in an intuitive way using a dynamically generated Birds's Eye View image. Furthermore, a novel classifier architecture to efficiently handle this complex application scenario is provided. By combining the processing speed of a classifier cascade with the discriminative power of a multi-stage neural network, the system achieves state of the art performance while retaining real-time capability. To keep classifier complexity low, a new feature-based inter-stage information transfer method is presented. All classifier components are compared to recent pedestrian detection approaches and evaluated on a real-world data set.