Fast Vehicle and Pedestrian Detection Using 2D Shapes and 3D Sizes

Abstract

Most existing pedestrian and vehicle detection algorithms use 2D features of objects, such as pixel values, color and texture, shape information or motion. The use of 3D features in object detection, on the other hand, are not well studied. In this paper, we propose a two-stage algorithm that uses both 2D and 3D features to detect pedestrian and vehicle in videos. The first stage compares the 2D contours between moving foreground and a set of object-specific shape kernels, to find the object candidates. A shape kernel is the contour of the object viewed from a certain panning and tilting angles, as well as a bounding cube that enclose this object. We have prepared 1584 and 1188 shape kernels for pedestrian and vehicle, respectively, and a speedup scheme is proposed to reduce the number of contour matching needed. The second stage further verifies the object candidates based on their 3D sizes, its width, height and length. Given shape kernels, the 3D sizes of the objects in a static scene monitored by a camera can be obtained using the intrinsic and extrinsic parameters of that camera. We employ a calibration-free method to estimate the camera parameters. The proposed algorithm can also handle partially occluded objects. Our experimental studies demonstrate that the proposed method can detect objects accurately and efficiently. It achieves a precision of 90%(90%) in detecting pedestrians(vehicles) at 15 frame-per-second.