Proposal of a Computer Vision System to Detect and Track Vehicles in Real Time Using an Embedded Platform Enabled with a Graphical Processing Unit

This paper presents a development proposal of a systemcapable of detecting, localizing and tracking vehicles inreal time using computer vision algorithms in an embeddedplatform for an advanced driver assistance system (ADAS). This project will use a Flea3 monocular color camera fromPointGrey. The video will be processed using support vectormachines (SVM) and convolutional neural networks (CNN). These algorithms will be implemented in a Jetson TK1 developmentboard from Nvidia which has a 192-core KeplerGPU, a quad-core ARM Cortex A15, 2 GB of RAM, 16 GBof internal memory, and a set of basic peripherals for automotiveapplications. Finally, this project will use librariessuch as: CUDA, OpenCV, and CudNN which are optimizedfor the Jetson TK1 as well as the LibSVM and Caffe librariesto train SVM and CNN models.

1. Introduction

Cars are one of the most frequently used forms of transport today, and they contribute to problems such as traffic jams, accidents and pollution in cities. From January 2014 to January 2015, there were 4769 deaths in car accidents in Colombia according to the DANE (National Administrative Department of Statistics) [6]. Now, traffic accidents are considered as a public health problem by the World Health Organization (WHO) [17].