DESIGN AND IMPLEMENTATION OF IMAGE-BASED OBJECT DETECTION AND TRACKING SYSTEM ON AUTONOMOUS SURVEILLANCE SMART CITY ROBOT

• Main Author: LUMBAN GAOL NIM : 13214004, GUNAWAN
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/27596
• Institution: Institut Teknologi Bandung

<p align="justify">This document addresses the embedded implementation problem of RGB based people detection and tracking from the perspective of moving observer. We focus on indoor scenarios in which robot is patrolling an area and take tracking action on supposedly suspicious people. Patrolling used point to point movement in 2D world coordinate. During patrolling, robot can take observation between each point to determine suspicious people. We define suspicious as to be in a relatively same location for an amount of duration. After exceeding timeout, robot will take tracking action until it is stopped or lose track of person. Losing track is defined after a certain amount of image frame processed without detecting any person. This document also describes findings on implementation of person detection and tracking algorithm in Nvidia Jetson TK1 using optimzed version of OpenCV for Jetson TK1 which is OpenCV4Tegra. Programming languanges used are python and C++. Main program uses python while functions are implemented using C++. This C++ function is called on main program by interfacing between C++ and python using Cython. We use built-in OpenCV HOG (Histogram of Oriented Gradients) descriptor for detection and KCF (Kernelized Correlation Filters) for tracking. The parameter values are then tuned to the best trade off values between speed and accuracy. Both detection and tracking algorithm is evaluated on data sample taken while robot is in function, providing realistic measurement on such real cases.<p align="justify"> <br /> <br />