Daniel's research website

About me

Name: Daniel Hu

School: Electrical Engineering at Fu Jen Catholic University, advised by Prof. Yuan-Kai Wang.

Lab: Intellgent System Lab

Research: Computer Vision and Deep Learning，Multiple Object Tracking

Email: daniel@islab.tw

Student : Chi-En Hu Advisor : Yuan-Kai Wang, Ph.D.

Department of Electrical Engineering

Fu Jen Catholic University

ABSTRACT

Online multiple object tracking is a good solution for many applications, such as autonomous driving and intelligence surveillance systems. For online multiple object tracking method, how to identify and associate the same target and determine the number of target in each frame are two core issues. With the development of deep learning, the application of deep feature model has become good solution to the problem of association. However, non-fixed object and non-fixed environment are still general challenges that feature models must solve. Our method uses the Single Task Joint Learning(STJL) model to the online multi-object tracking framework, and STJL model can learn training data from different environments and improve generalization. On the other hand, we propose initialization judgment to solve the problem of determining the number of target in each frame. The experiment showed that STJL model can improve accuracy in all test environment, and by STJL model and proposed tracker initialization method, our MOT method can obtain well MOT accuracy in many different environments.

第17屆離島資訊技術與應用研討會

2018 Conference on Information Technology and Applications in Outlying Islands

最佳論文獎

胡啟恩、郭雅諒、王元凱

輔仁大學電機工程學系

摘要

目前大多數多物件追蹤的方法都採用Tracking-by-detection的架構，這樣的架構大致可以分為二個部分，分別是物件偵測器與追蹤器。由於此架構需要先仰賴物件偵測器來確定物件的存在與位置，接著才會做後續追蹤動作，因此物件偵測器的效果會直接影響到最後的追蹤結果。本論文實現了一個多物件追蹤系統，在多物件追蹤中結合了一個高準確率的偵測器，以改善偵測器的方式改善最後的追蹤結果。我們使用一個深度學習的方法Multi-scale CNN作為系統的偵測器，而追蹤器則使用MDP Track追蹤演算法。實驗以PET2009與TUD-Crossing資料集進行驗證，實驗結果顯示，比起原方法及其他先進的追蹤方法，我們的方法有著更佳的追蹤正確率。

科儀新知215期，第36-47頁，2018年6月

胡啟恩、郭雅諒、王元凱

輔仁大學電機工程學系

摘要

現今人工智慧受到高度矚目與快速發展，電腦視覺身為其中之關鍵技術也已成為許多研究及應用的焦點，而隨著智慧型視訊監控與無人駕駛自動車的快速發展，以電腦視覺方法進行影像視訊中的行人偵測與追蹤，更成為一個重要而關鍵的技術。行人追蹤與行人偵測在電腦視覺研究上為兩項不同的技術，但是兩者在應用面卻有十分緊密的關係。近幾年因深度學習的崛起，行人偵測正確率有顯著的進步，也使得行人追蹤有更好的效果。本文將詳細介紹行人偵測追蹤技術，包括行人偵測與行人追蹤技術的發展，並介紹在線多物件追蹤系統之技術內容。

Abstract

As of today, Artificial Intelligence has gained a lot of attention and is currently being developed rapidly. As one of the key technologies, Computer Vision has become the focal point of many researches and applications, due to the rapid development of pedestrian detection and tracking methods used in Intelligent Surveillance Systems and Self-driving Cars. Pedestrian detection and tracking are two different technologies, yet inseparable when used in applications. Due to the rise of deep learning, pedestrian detection has made great progress in accuracy, which also has driven the improvement of pedestrian tracking. In this paper, we in detail go through the introduction of pedestrian detection and tracking technologies, including the development of both technologies, and also the introduction of online multiple object tracking systems.