Tracking People by Detection Using CNN Features

Abstract

Multiple people tracking is an important task for surveillance. Recently, tracking by detection methods had emerged as immediate effect of deep learning remarkable achievements in object detection. In this paper, we use Faster-RCNN for detection and compare two methods for object association. The first method is simple Euclidean distance and the second is more complicated Siamese neural network. The experiment result show that simple Euclidean distance gives promising result as object association method, but it depends heavily on the robustness of detection process on individual frames.