Gpu architecture for stationary multisensor pedestrian detection at smart intersections

Abstract

We present a real-time multisensor architecture for combined laser scanner and infra-red video-based pedestrian detection and tracking used within a road side unit for intersection assistance. In order to achieve outmost classification performance we propose a cascaded classifier using laser scanner hypothesis generation and histogram of oriented gradients (HOG) descriptors for video-based classification together with linear and Gaussian kernel support vector machines (SVM). The entire classification cascade is implemented on a graphics processing unit (GPU). Giving real-time performance top priority, we present novel compute unified device architecture (CUDA) implementations of a selective HOG-based feature extraction and background subtraction based on mixture of Gaussians (MOG). The classification cascade is managed by a multi-core CPU that further performs pedestrian tracking using a linear Kalman filter. Evaluation on an infra-red benchmark database and an experimental study on a real-world intersection used within the Ko-PER project confirm excellent classification and real-time performance around the clock without external illumination.