Estudo de métodos de inteligência computacional para detecção de humanos em imagens de webcam

Abstract

This undergraduate thesis presents the implementation of two different human detection algorithms, one using a Matlab tool that uses histograms of oriented gradients feeding a support vector machine, and other using a deep convolutional neural network. Basic concepts that are necessary to comprehend the methods, like neural network definitions, convolutions and gradient calculation, are presented. The basic structure of this thesis can be divided into two parts: Peopledetector, utilized on the first detection method, and Mobilenet-SSD, utilized on the second one. The necessary concepts to the comprehension of the detection tool and its parameters are presented on the first part. On the second part, the Caffe architecture and the Mobilenet and Single Shot Multibox Detector models are explained in more details. The first method is implemented on Matlab while the second algorithm is programmed in Python using the OpenCV and Imutils libraries. The PeopleDetector works better on situations that involves humans standing in upright positions, and even then, the method presents some false positive detections. The deep convolutional neural network presented excellent results in precision and speed, showing as a possible option for embedded applications.