Below are some examples of previous FYP oral presentations and FYT oral defences that may be helpful when you prepare yours.
| 2020 LIX1 Smart Parking Lot with Real-time Space Monitoring |
This project leveraged technologies like the Internet of Things, cloud computing, and artificial intelligence to help the operations of the HKUST car park could be smoothened and speeded up with a reduction in human resources required. In this project, the team aimed to develop an integrated solution to digitize car park operations, such as visitor guidance, payment, access control, and park management. The result was a data-driven, integrated smart car park system that featured a wide range of components, ranging from sensor devices, cloud-based backend, a tablet app for gate kiosks, a mobile app for drivers, to a web management portal. Data collection and analysis tasks like live telemetry collection and license plate recognition were continuously carried out. Utilizing the real-time data collected, the system then provided useful features to visitors and staff of the car park, such as visualization of lot vacancy, self-serviced e-payment, automated access control, visitor management, and car park usage analytics. |
| 2020 LIX2 USmarTpark: Smart Parking Lot with Real-time Space Monitoring |
This project utilizes the internet of things (IoT) technology to capture and collect images of parking spaces, artificial intelligent (AI) approach to detect the occupancy status and extract the vehicle license plate number and cross-platform (Android and iOS) native mobile application to display the occupancy information in real-time. The IoT devices were reliable and were tested in early March 2020 in the HKUST indoor car park at LG2. With the help of SSC and CMO, testing was conducted smoothly in a twenty-four hour manner. The stability and reliability of the IoT devices were well proven, and the design can be used in the long term. |
| 2020 LZ1 Liveness Detection in Facial Recognition for a Security Gateway |
In order to detect whether people's skin is real or not, spectroscopic cameras with infrared light can be used to analyze the liveness of their faces. The basic theory of this project was that the reflection of lights on different materials differs significantly. Both RGB and NIR (near-infrared) algorithms were applied in this project. An RGB algorithm was used for detecting flesh color, while an NIR algorithm was used for detecting human skin. With the combination of both algorithms, only real human faces are detected. Three spectroscopic cameras with different filters were used to capture NIR light, while a normal digital camera was used to capture visible light. Lighting equipment that emits different wavelengths of lights was also applied to enhance the reflection. Hyperspectral data, as well as three-channel RGB image data, was collected and processed to reach the results. The project was sponsored and supported by Yitu Technology |
| 2020 LZ2 Smart Traffic Signals Using Deep Learning-based Computer Vision |
This project attempted to utilize deep learning-based computer vision to develop smart traffic signals that can adapt to the current traffic conditions and optimize green duration for each lane in an intersection, resulting in a reduction of the average waiting time of vehicles and improvements in the traffic efficiency. |
| 2020 RICWTAO1 Machine Learning to Detect and Identify Web Application Attacks from Big Web Server Log Data |
Some web applications do not have sufficient time to undergo comprehensive reviews and tests. This creates a good chance for attackers to look for lucrative vulnerabilities, as well as sensitive info. However, analyzing weblogs consumes lots of cyber professionals' manpower, and the task is repetitive. To counter the threat effectively, there is an urge and demand to detect known and novel attacks through machine learning techniques. In this project, machine learning algorithms and techniques were applied to (1) to detect web application attacks, (2) to analyze the success of web application attacks, and (3) to distinguish the type of web application attacks. Using knowledge in penetration testing, the team generated and collected multiple datasets for training and testing machine learning models. After various stages of machine learning, they found the results for (1) and (3) to be satisfying. |
| 2020 SYQ1 Extracting Relationships among Entities from Newspaper Articles |
In this project, the team strived to develop an automated method for relation extraction from news articles by borrowing ideas from Natural Language Processing (NLP). They restructured the problem as a sequence of multiple NLP tasks, including Coreference Resolution, Named Entity Recognition, Relation Extraction and Sentiment Analysis. They utilized state-of-the-art deep learning models and frameworks for each of them. They used few shot relation extraction to do the relation extraction step, and they encapsulated their automation in a web application with an intuitive user interface that met the use-case requirements of target users. Finally, they put some additional annotation, SNA and visualization tools at the users' disposal to allow them to customize and interact with the platform. The system has some potential to allow for much faster relation extraction in the Political Science field. |
| 2018 RO4 Optimal Investment Strategy Using Scalable Machine Learning and Data Analytics for Small-cap Stocks |
Small capitalization stocks are characterised by higher volatility and higher potential returns. This project provided a platform for small cap stock trading. It included price prediction (based on multiple regression), portfolio allocation (based on user inputs and a convex optimization library based on Markowitz’s Mean Variance Theorem) and a web application. (Distinguished project) |
| 2015 LZ1 Topic-Based Browsing of the Amazon Discussions Feedback Forum |
Utilizes latent Dirichlet allocation to automatically organize the posts on an Internet forum based on the topics the posts; displays the results on a website so that users can directly browse the forum by topic |
| 2015 MA1 Leap Sense: Turn Any Computer Screen into a Gesture-Assisted Touch Screen Using a Leap Motion Controller |
Utilizes a Leap Motion Controller to detect hand movement and enable gesture control and virtual touch screen functionality; distinguished project; 2015 HKUST President's Cup Winner! |
| 2020 PAN1 Age of Robot Warlords: an Augmented Reality Mobile Strategy Game for Multiple Players |
This game is an AR mobile strategy game for multiple players called the Age of Robot Warlords. It was developed with the Unity3D game engine, C#, Blender, the Photon Unity Networking 2 API and Vuforia. The AR allows players to interact with real environments to make their strategies. |
| 2015 PAN4 Presentation Tools with Gesture Recognition and Augmented Reality |
Cool and dynamic presentation software using a Leap Motion controller and adding Augmented Reality (AR) into the presentations; includes an add-in for MS PowerPoint (Distinguished project) |
| 2020 WIL1 A Blockchain Database Application for Connecting Domestic Helpers and Hong Kong Employers |
This project explored the possibility of building a platform for connecting domestic helpers and Hong Kong employers, backed by a blockchain database. The value we introduced is the theoretical resistance to anyone who would like to dishonestly tamper with the records in the database, including domestic helpers, employers themselves or agencies. |
| 2020 KWT2 Petriversity - Gamification in Education |
This system is a real-time learning game platform for learning in a university. Players are able to create questions and wait for their classmates to rate the questions. Rated questions appear when players initiate combat with AI enemies. They have to answer a question in order to defeat the enemy and complete their time in a dungeon to receive virtual or real-life rewards. The system also includes other gamification features, such as a leaderboard and an achievement board, to promote the learning experience. |
| 2020 KWT3 Wireless VR Game with Gesture and Voice Recognition |
This VR game integrates with a wireless VR headset, IBM Watson, Microsoft Speech Platform, Leap Motion, an accelerometer and fan to free the player from the restraints of traditional devices and allow more "reality" to be put into the game. |
| 2020 YIKE2 First-Person Computer Shooter Game with Eye Tracking and Unity |
FPS computer game which incorporates eye-tracking technology as an alternative of mouse game input; developed in Unity using C# with the Tobii Unity software development kit. (Distinguished Project) |
| 2013 RO2 Battle Fury: Multiplayer Game |
A prototype third-person action multiplayer game developed with Unreal Script, Maya, 3D Studio Max, Mudbox, World Machine, GIMP, Photoshop and the Unreal Development Kit |
| 2015 PSAN2 Crowbar: 3D Level 'Map' Editor for Video Games |
A new editor intended to help streamline the development process of a 3D computer game level, maintain user focus on abstract tasks that are relevant to the design process, automate the subsequent operations required to achieve these tasks and assist users in constructing complex geometry |
| 2020 LZ1 Liveness Detection in Facial Recognition for a Security Gateway |
In order to detect whether people's skin is real or not, spectroscopic cameras with infrared light can be used to analyze the liveness of their faces. The basic theory of this project was that the reflection of lights on different materials differs significantly. Both RGB and NIR (near-infrared) algorithms were applied in this project. An RGB algorithm was used for detecting flesh color, while an NIR algorithm was used for detecting human skin. With the combination of both algorithms, only real human faces are detected. Three spectroscopic cameras with different filters were used to capture NIR light, while a normal digital camera was used to capture visible light. Lighting equipment that emits different wavelengths of lights was also applied to enhance the reflection. Hyperspectral data, as well as three-channel RGB image data, was collected and processed to reach the results. The project was sponsored and supported by Yitu Technology |
| 2017 RAYW2 Automatic Parking Space Allocation and Indoor Parking Lot Navigation System with Beacons |
The result of this project was a prototype for a next-generation car parking system, which provides automatic parking space allocation as well as indoor navigation, to improve car parking experience. Inspired by the concept of sharing economy, parking space sharing was also included to enable sharing of unused parking spaces so as to resolve the problem of insufficient parking space during peak hours. The project was sponsored by Radica, a digital marketing company located in the Hong Kong Science Park. The systme includes two native mobile applications, one for iOS using Swift 3 while another one for Android using Java, integrated with cloud mobile backend on Amazon EC2 to provide real-time parking space vacancy data, one-click car park check-in, and check-out with ePayment. The indoor positioning is powered by beacons, which are small Bluetooth Low Energy (BLE) devices that are popularly used for proximity marketing and indoor positioning. |
| 2013 RO3 Stock Data Market Analysis System |
A tool for an investment company, HWK, that uses pre-processing and pattern recognition algorithms to automate and optimize the price patter recognition process for six common stock chart patterns |
| 2020 RIC2 Security Compliance Checker Tools for Cloud Environments v2 |
The security settings of cloud services are often more complicated than local machines, so users may not be able to verify and secure their cloud environment settings with ease. The aim of this project was to improve the security compliance checker tools developed by last year’s FYP group. It provides clear visuals and interactive elements, such as network maps, so that users are able to examine and discover hidden insights and security findings on their cloud resources using the updated security checker tools. Various new features implemented in the tools also enable users to customize their assessment reports and construct self-defined security rules for extending the usability of these tools in an actual scenario. |
| 2020 RICWTAO1 Machine Learning to Detect and Identify Web Application Attacks from Big Web Server Log Data |
Some web applications do not have sufficient time to undergo comprehensive reviews and tests. This creates a good chance for attackers to look for lucrative vulnerabilities, as well as sensitive info. However, analyzing weblogs consumes lots of cyber professionals' manpower, and the task is repetitive. To counter the threat effectively, there is an urge and demand to detect known and novel attacks through machine learning techniques. In this project, machine learning algorithms and techniques were applied to (1) to detect web application attacks, (2) to analyze the success of web application attacks, and (3) to distinguish the type of web application attacks. Using knowledge in penetration testing, the team generated and collected multiple datasets for training and testing machine learning models. After various stages of machine learning, they found the results for (1) and (3) to be satisfying. |
| 2020 RIC1 HUNT - Automatic Search Bar Finder and Search Time Saver for Hong Kong |
Search engine results nowadays are not always satisfactory, because the top results are usually being shaped by certain ranking algorithm, making some results listed prior to others. Therefore, this project focused on retrieving more relevant results by utilizing the internal search functions in selected websites. The system includes a mechanical user that can send keywords to search bars in a set of popular Hong Kong websites and perform searching automatically in those websites. This feature relies on data scraping algorithms for different types of search bars in websites. An executable program was also developed that enables users to use this application for any selected website. |
| 2016 LUO2 RPG Teaching Assistant Management System |
A website with a responsive UI to help the CSE Department professors and administrative staff assign TAs to courses each semester, including a matching program that takes into consideration a number factors, including course requirements and schedules, the preferences of advisors and the experience, research areas and preferences of PG students |
| 2016 RAYW1 A Mobile-friendly HKUST Facility Booking System |
Since the existing Facility Booking System (FBS) was designed and implmented many years ago before smartphones and mobile apps became popular, this new booking system is a web-based system that allows more user-friendly booking and administrative support as well as tighter security. It includes a facility booking page for the general users (i.e., student, society members, and staff ) to book facilities, and it has an admin page for administrative staff to manage the system (i.e., approve bookings and manage accounts, facilities and equipment.) |
| 2013 LUO2 A Web Portal for Hiking in HK |
Utilization of Google fusion tables and KML files for indicating itineraries on HK trails |
| 2020 KWT2 Petriversity - Gamification in Education |
This system is a real-time learning game platform for learning in a university. Players are able to create questions and wait for their classmates to rate the questions. Rated questions appear when players initiate combat with AI enemies. They have to answer a question in order to defeat the enemy and complete their time in a dungeon to receive virtual or real-life rewards. The system also includes other gamification features, such as a leaderboard and an achievement board, to promote the learning experience. |
| 2020 MA1 How Good Is Your English Pronunciation? A Computer-aided Pronunciation Training System Using Deep Neural Networks |
This is a mobile application that teaches users basic phonetics by utilizing a neural-network that is able evaluate user speech and give feedback on the phoneme level in real-time. It can also keep track of their performance over time. |
| 2020 NGOK3 An Automated English Vocabulary Learning Mobile Application |
This unique application provides different ways to tackle the issue of memorizing vocabulary. It allows users to read news articles based on their preferences, and it highlights vocab words help users memorize them. It also includes flashcards and graphics to enable easier memorization, as well as study mode to help users retaining the word knowledge for longer time. And it has a space repetition algorithm which calculates the forgetting time since a word was last reviewed. |
| 2016 LIX2 HKUST Hour of Code - iOS Mobile App to Teach Children Basic Scratch Programming Skills and Concepts |
A free and interactive iPad App for children to learn how to code with Scratch; includes a carefully-designed curriculum to teach basic programming fundamentals with well-known fairy tales, guided tutorials for kids to go through to pick up basic coding skills, project templates for free coding to help kids unleash their imagination, a special Hour of Code tutorial for a quick taste of coding within just an hour and beautiful graphics and the attractive user interface (Distinguished project) |
| 2020 MXJ2 Positive Response Recommendation in Social Media Based on Human Need Analysis |
This project aimed to develop a mapping from text analysis of social media posts to human needs and then develop a model that takes such needs into consideration to make recommendations for positive responses on seeing new posts. To do this, historic social media posts were first gathered and analyzed to understand the underlying concerns and needs of users. Need profiles were then formulated and utilized as a semantical context to understand new textual social media posts. This made it possible to generate keywords for compliments according to relevance towards both users' concerns and the content of target posts. Next, happy stories could be recommended with similar underlying needs to enrich the content of responses. A prototype application was built for the entire recommendation pipeline, and then the psychological impact of positive responses was analyzed on post owners. |
| 2020 RAYW2 When Does GCN Work: Understanding and Improving Neighborhood Aggregation |
Graph Neural Networks (GNNs) have been shown to be powerful in a wide range of graph-related tasks. While there exist various GNN models, a critical common ingredient is neighborhood aggregation, where the embedding of each node is updated by referring to the embedding of its neighbors. This thesis aimed to provide a better understanding of this mechanism by asking the following question: Is neighborhood aggregation always necessary and beneficial? In short, the answer is no. Two conditions under which neighborhood aggregation is not helpful are: (1) when the neighbors of a node are highly dissimilar and (2) when a node’s embedding is already similar to that of its neighbors. Some novel metrics are proposed that quantitatively measure these two circumstances and integrate them into an adaptive-layer module. Experiments show that allowing for node-specific aggregation degrees has significant advantage over current GNNs. |
| 2020 SC1 Vulnerability Analysis of Deep Neural Networks |
This thesis carefully observes how adversarial samples fool DNN models to misclassify within MNIST, CIFAR10, and Tiny ImageNet. The designed defense mechanism could generally protect against a broad range of attacks under the experiments, and the results were slightly better than some state-of-the-art techniques, like Neural Network Invariant Checking (NIC) and Local Intrinsic Dimensionality (LID), for detecting adversarial attacks under 7 adversarial attacks and 12 commonly used DNN models. |
| 2017 RAYW3 RAYW3 Finding Maximum Regret Ratio Minimizing Sets of Minimized Size |
Selecting a subset of points from a database that are representative of the information in the database is a problem with many applications, from providing suggestions to users to summarizing the information in the database. Looking at the problem of selecting such a representative set from the perspective of the users, it is important to include the datapoints in the set that ensure that all the users are satisfied with the set, at least to a certain extent. This thesis discusses how the satisfaction of the users can be quantified, and an algorithm is proposed by which a subset of the database can be selected that guarantees that all the users are satisfied with the selected points to a desired extent. This algorithm is shown to be an O(dlogn)-approximation algorithm, where d and n are the dimensionality and size of the database respectively. Furthermore, the problem addressed is shown to be NP-Hard. |
| 2016 RAYW3 Enhanced Heuristic Algorithms for Seed Set Selection in Influence Maximisation |
Using the lowest cost to generate the largest influence has long been an important strategy to advertise products. With the help of social networks, information diffusion has become more effective on the Internet. The selection of seed users helps in maximising the influence of product advertising in social networks. This thesis shows that using heuristic algorithms leads to a reasonable size of influence using the shortest time and the lowest memory consumption. Two new heuristic algorithms are proposed, Mixed Degree Influence Discount and Second Level Degree Influence Discount, in which each vertice's score is a combination of its degree and its outgoing edge weight. The performance was compared to that of other algorithms by carrying out experiments using datasets of different sizes. |
| 2020 RO3 New Trading Strategies Inspired from Harmonic and Elliott Wave Trading Strategies |
The Harmonic Trading Strategy and related Elliott Wave Theory have led to revolutions in the field of Forex trading. The team implemented both strategies on the Quantopian platform and found that neither strategy is remotely as powerful when applied to other asset classes, particularly equities. They fall victim to the highly volatile market noise, often ending up losing more money than hands-off passive strategies. Thus, they tried a new strategy that incorporates both Harmonic and Elliott patterns, but is specialized for trading in the equities market, as it can adapt to uncertain market fluctuations. | Final Report |
| 2018 RO4 Optimal Investment Strategy Using Scalable Machine Learning and Data Analytics for Small-cap Stocks |
Small capitalization stocks are characterised by higher volatility and higher potential returns. This project provided a platform for small cap stock trading. It included price prediction (based on multiple regression), portfolio allocation (based on user inputs and a convex optimization library based on Markowitz’s Mean Variance Theorem) and a web application. (Distinguished project) |
| 2020 LIX1 USmarTpark: Smart Parking Lot with Real-time Space Monitoring |
This project utilizes the internet of things (IoT) technology to capture and collect images of parking spaces, artificial intelligent (AI) approach to detect the occupancy status and extract the vehicle license plate number and cross-platform (Android and iOS) native mobile application to display the occupancy information in real-time. The IoT devices were reliable and were tested in early March 2020 in the HKUST indoor car park at LG2. With the help of SSC and CMO, testing was conducted smoothly in a twenty-four hour manner. The stability and reliability of the IoT devices were well proven, and the design can be used in the long term. |
| 2020 LIX2 Smart Parking Lot with Real-time Space Monitoring |
This project leveraged technologies like the Internet of Things, cloud computing, and artificial intelligence to help the operations of the HKUST car park could be smoothened and speeded up with a reduction in human resources required. In this project, the team aimed to develop an integrated solution to digitize car park operations, such as visitor guidance, payment, access control, and park management. The result was a data-driven, integrated smart car park system that featured a wide range of components, ranging from sensor devices, cloud-based backend, a tablet app for gate kiosks, a mobile app for drivers, to a web management portal. Data collection and analysis tasks like live telemetry collection and license plate recognition were continuously carried out. Utilizing the real-time data collected, the system then provided useful features to visitors and staff of the car park, such as visualization of lot vacancy, self-serviced e-payment, automated access control, visitor management, and car park usage analytics. |
| 2020 NGOK1 A HKUST Restaurant Recommendation System |
In this project, the team built a restaurant recommendation system for HKUST, where the recommendations are based on in-campus restaurants. The system utilizes machine learning with computer vision technology and real-time navigation through an indoor positioning system to provide users with suitable restaurant recommendations. It includes a cross-platform application and a server module. |
| 2020 NGOK2 Smart Personal Helper for HKUST Students |
This Android mobile application is a smart personal helper that integrates different schedulers for HKUST students and applies a deep learning technique to recommend suitable events to different users. It also provides a calendar, course deadline management function and a groupmate finding function. |
| 2020 RAYW1 UST Path Advisor System 2.0: Image-based Localization Using a CNN |
This project enhanced the HKUST Path Advisor by adding a location estimation feature based on a computer vision technique that relies on a Convolutional Neural Network (CNN) and a feature matching algorithm. Upon receiving a query image uploaded by the user, the model executes multi-class image classification in real-time with efficiency and returns the location estimation to the user. The core achievement of this project was eliminating the need to install or utilize wireless network infrastructures like a signal-based or RFID receivers, which are commonly used in other popular indoor positioning systems. |
| 2016 RAYW1 A Mobile-friendly HKUST Facility Booking System |
Since the existing Facility Booking System (FBS) was designed and implmented many years ago before smartphones and mobile apps became popular, this new booking system is a web-based system that allows more user-friendly booking and administrative support as well as tighter security. It includes a facility booking page for the general users (i.e., student, society members, and staff ) to book facilities, and it has an admin page for administrative staff to manage the system (i.e., approve bookings and manage accounts, facilities and equipment.) |
| 2020 MA1 How Good Is Your English Pronunciation? A Computer-aided Pronunciation Training System Using Deep Neural Networks |
This is a mobile application that teaches users basic phonetics by utilizing a neural-network that is able evaluate user speech and give feedback on the phoneme level in real-time. It can also keep track of their performance over time. |
| 2020 MA2 CSE Chatbot |
In this project, the team developed a chatbot application on Android and iOS platforms, to provide meaningful responses and guidance on users' questions related to the HKUST CSE department webpage. The Chatbot could also engage in some small talk with the users if their inputs are irrelevant to the CSE website. The system is a combination of a conversational and task-oriented chatbot, and it was built with hand-crafted dialog management, a natural language understanding algorithm and a simple machine learning-based algorithm. |
| 2020 MA3 An Android App for Cantonese Lipreading Using the LipNet Technology |
This project extended the field of automated lipreading to Cantonese with a newly developed machine learning technology, LipNet. A Cantonese LipNet model was trained with data from 45 speakers speaking 20 commonly used Cantonese phrases. The experiments resulted in an accuracy of 93.4% in detecting the correct phrases. The Cantonese lipreading model was packaged in a user-friendly Android mobile application for convenient instant lipreading and exploration of potential applications. |
| 2020 MXJ1 Meaningful Metaphor Mining: Identification and Explanation |
In this project, the team conducted plentiful research on a specific domain: metaphor understanding. To understand the meaningfulness of metaphors, they collected 6,104 human-labeled meaningful pairs from large corpora, and 1295 among them were labeled as meaningful. After doing feature analysis, they collected 765 dimensions of features in three different feature categories: psycholinguistic, conceptual, and SynSet features. Then, they adopted LightGBM to complete this metaphor meaningfulness identification work as a binary classification task. After some human-in-the-loop validation, they derived an identification model with AUC up to .895. Next, they implemented three unsupervised methods to explain a metaphor word pair: a heuristic model, a BERT masked model and a fusion model. Then, they conducted a final user study to systematically evaluate these explanation models. |
| 2020 SYQ1 Extracting Relationships among Entities from Newspaper Articles |
In this project, the team strived to develop an automated method for relation extraction from news articles by borrowing ideas from Natural Language Processing (NLP). They restructured the problem as a sequence of multiple NLP tasks, including Coreference Resolution, Named Entity Recognition, Relation Extraction and Sentiment Analysis. They utilized state-of-the-art deep learning models and frameworks for each of them. They used few shot relation extraction to do the relation extraction step, and they encapsulated their automation in a web application with an intuitive user interface that met the use-case requirements of target users. Finally, they put some additional annotation, SNA and visualization tools at the users' disposal to allow them to customize and interact with the platform. The system has some potential to allow for much faster relation extraction in the Political Science field. |
| 2020 MXJ1 Meaningful Metaphor Mining: Identification and Explanation |
In this project, the team conducted plentiful research on a specific domain: metaphor understanding. To understand the meaningfulness of metaphors, they collected 6,104 human-labeled meaningful pairs from large corpora, and 1295 among them were labeled as meaningful. After doing feature analysis, they collected 765 dimensions of features in three different feature categories: psycholinguistic, conceptual, and SynSet features. Then, they adopted LightGBM to complete this metaphor meaningfulness identification work as a binary classification task. After some human-in-the-loop validation, they derived an identification model with AUC up to .895. Next, they implemented three unsupervised methods to explain a metaphor word pair: a heuristic model, a BERT masked model and a fusion model. Then, they conducted a final user study to systematically evaluate these explanation models. |
| 2020 MXJ2 Positive Response Recommendation in Social Media Based on Human Need Analysis |
This project aimed to develop a mapping from text analysis of social media posts to human needs and then develop a model that takes such needs into consideration to make recommendations for positive responses on seeing new posts. To do this, historic social media posts were first gathered and analyzed to understand the underlying concerns and needs of users. Need profiles were then formulated and utilized as a semantical context to understand new textual social media posts. This made it possible to generate keywords for compliments according to relevance towards both users' concerns and the content of target posts. Next, happy stories could be recommended with similar underlying needs to enrich the content of responses. A prototype application was built for the entire recommendation pipeline, and then the psychological impact of positive responses was analyzed on post owners. |
| 2016 PAN3 Real-time Emotion Sensing with Google Glass |
An emotion sensing system with Google Glass, which allows users to infer the emotions of others by analyzing their facial expressions in an automated yet unobtrusive way. The inferring process is based on machine learning and support vector machine (SVM) emotion classifiers that represent six plus one distinct emotions according to Ekman’s basic emotion model. These classifiers were painstakingly devised by first manually selecting 1,806 representative images of Asian faces from 5553 images in eight different open source facial expression datasets. Then, through OpenCV’s face detector and dlib’s face landmark extractor, 18 key landmark points were identified on each face and 17 displacement vector features were calculated for each of the seven emotions. These features are fed into LIBSVM’s machine learning library for training and emotion prediction. The combined system enables real-time emotion sensing on the Google Glass with good accuracy. The levels of sensed emotions are overlaid with augmented reality (AR) on the Android mobile phone user interface as well as audibly conveyed via the tiny headset speaker in the Google Glass platform. |
| 2020 LIX1 USmarTpark: Smart Parking Lot with Real-time Space Monitoring |
This project utilizes the internet of things (IoT) technology to capture and collect images of parking spaces, artificial intelligent (AI) approach to detect the occupancy status and extract the vehicle license plate number and cross-platform (Android and iOS) native mobile application to display the occupancy information in real-time. The IoT devices were reliable and were tested in early March 2020 in the HKUST indoor car park at LG2. With the help of SSC and CMO, testing was conducted smoothly in a twenty-four hour manner. The stability and reliability of the IoT devices were well proven, and the design can be used in the long term. |
| 2020 LIX2 Smart Parking Lot with Real-time Space Monitoring |
This project leveraged technologies like the Internet of Things, cloud computing, and artificial intelligence to help the operations of the HKUST car park could be smoothened and speeded up with a reduction in human resources required. In this project, the team aimed to develop an integrated solution to digitize car park operations, such as visitor guidance, payment, access control, and park management. The result was a data-driven, integrated smart car park system that featured a wide range of components, ranging from sensor devices, cloud-based backend, a tablet app for gate kiosks, a mobile app for drivers, to a web management portal. Data collection and analysis tasks like live telemetry collection and license plate recognition were continuously carried out. Utilizing the real-time data collected, the system then provided useful features to visitors and staff of the car park, such as visualization of lot vacancy, self-serviced e-payment, automated access control, visitor management, and car park usage analytics. |
| 2020 RAYW1 UST Path Advisor System 2.0: Image-based Localization Using a CNN |
This project enhanced the HKUST Path Advisor by adding a location estimation feature based on a computer vision technique that relies on a Convolutional Neural Network (CNN) and a feature matching algorithm. Upon receiving a query image uploaded by the user, the model executes multi-class image classification in real-time with efficiency and returns the location estimation to the user. The core achievement of this project was eliminating the need to install or utilize wireless network infrastructures like a signal-based or RFID receivers, which are commonly used in other popular indoor positioning systems. |
| 2017 RAYW2 Automatic Parking Space Allocation and Indoor Parking Lot Navigation System with Beacons |
The result of this project was a prototype for a next-generation car parking system, which provides automatic parking space allocation as well as indoor navigation, to improve car parking experience. Inspired by the concept of sharing economy, parking space sharing was also included to enable sharing of unused parking spaces so as to resolve the problem of insufficient parking space during peak hours. The project was sponsored by Radica, a digital marketing company located in the Hong Kong Science Park. The systme includes two native mobile applications, one for iOS using Swift 3 while another one for Android using Java, integrated with cloud mobile backend on Amazon EC2 to provide real-time parking space vacancy data, one-click car park check-in, and check-out with ePayment. The indoor positioning is powered by beacons, which are small Bluetooth Low Energy (BLE) devices that are popularly used for proximity marketing and indoor positioning. |
| 2020 KWT1 Smart Scheduler |
In this project, the team developed a mobile app called 'Smart Scheduler' to increase the efficiency of completing regular calendar tasks and to help the user improve his or her time management. The app is like a time tracker with auto-completion functionality. Utilizing machine learning on big data from a large group of users, the app can learn the behavior of each user and track how he or she uses his or her time. Then it can make suggestions to help users schedule events and save time. The app also automates some common tasks, like finding traffic conditions, predicting the quickest path to the next event, suggesting activities for free time slots and automatically filling in free time slots. It also allows users to create group schedules and checklists. |
| 2020 MA2 CSE Chatbot |
In this project, the team developed a chatbot application on Android and iOS platforms, to provide meaningful responses and guidance on users' questions related to the HKUST CSE department webpage. The Chatbot could also engage in some small talk with the users if their inputs are irrelevant to the CSE website. The system is a combination of a conversational and task-oriented chatbot, and it was built with hand-crafted dialog management, a natural language understanding algorithm and a simple machine learning-based algorithm. |
| 2020 MA3 An Android App for Cantonese Lipreading Using the LipNet Technology |
This project extended the field of automated lipreading to Cantonese with a newly developed machine learning technology, LipNet. A Cantonese LipNet model was trained with data from 45 speakers speaking 20 commonly used Cantonese phrases. The experiments resulted in an accuracy of 93.4% in detecting the correct phrases. The Cantonese lipreading model was packaged in a user-friendly Android mobile application for convenient instant lipreading and exploration of potential applications. |
| 2020 NGOK1 A HKUST Restaurant Recommendation System |
In this project, the team built a restaurant recommendation system for HKUST, where the recommendations are based on in-campus restaurants. The system utilizes machine learning with computer vision technology and real-time navigation through an indoor positioning system to provide users with suitable restaurant recommendations. It includes a cross-platform application and a server module. |
| 2020 NGOK2 Smart Personal Helper for HKUST Students |
This Android mobile application is a smart personal helper that integrates different schedulers for HKUST students and applies a deep learning technique to recommend suitable events to different users. It also provides a calendar, course deadline management function and a groupmate finding function. |
| 2020 NGOK3 An Automated English Vocabulary Learning Mobile Application |
This unique application provides different ways to tackle the issue of memorizing vocabulary. It allows users to read news articles based on their preferences, and it highlights vocab words help users memorize them. It also includes flashcards and graphics to enable easier memorization, as well as study mode to help users retaining the word knowledge for longer time. And it has a space repetition algorithm which calculates the forgetting time since a word was last reviewed. |
| 2016 LIX2 HKUST Hour of Code - iOS Mobile App to Teach Children Basic Scratch Programming Skills and Concepts |
A free and interactive iPad App for children to learn how to code with Scratch; includes a carefully-designed curriculum to teach basic programming fundamentals with well-known fairy tales, guided tutorials for kids to go through to pick up basic coding skills, project templates for free coding to help kids unleash their imagination, a special Hour of Code tutorial for a quick taste of coding within just an hour and beautiful graphics and the attractive user interface (Distinguished project) |
| 2013 LZ2 A TCM Advisory Android App for Sub-Health (II) |
This Android app diagnoses people’s health condition by asking them questions about different symptoms and provides health suggestions by using 2000 sample data collected in Beijing by TCM professionals. |
| 2012 LUO1 iCDD - An iPhone-based Choh Dai Di Card Game |
iPhone Choi Dai Di game with the popular initial card exchange; includes beginning, intermediate and advanced AI levels of difficulty; good introduction for Cho Dai Di beginners; developed with Xcode with the cocos2d plug-in |
| 2020 KWT3 Wireless VR Game with Gesture and Voice Recognition |
This VR game integrates with a wireless VR headset, IBM Watson, Microsoft Speech Platform, Leap Motion, an accelerometer and fan to free the player from the restraints of traditional devices and allow more "reality" to be put into the game. |
| 2016 MA2 Using the Leap Motion Controller to Translate Sign Language to Speech |
This system returns speech and text when the user performs sign language in front of a Leap Motion Controller (LMC). The LMC is used to capture hand gesture images and convert them into the positional and direction information. This data is compared with the data inside the database to determine the most similar sign using a tailor-made Dynamic Time Warping (DTW) algorithm. Once a gesture has been recognized, its corresponding meaning will be presented in both audio and text format. (Distinguished project; 2016 HKUST President's Cup Gold Award!) |
| 2015 MA1 Leap Sense: Turn Any Computer Screen into a Gesture-Assisted Touch Screen Using a Leap Motion Controller |
Utilizes a Leap Motion Controller to detect hand movement and enable gesture control and virtual touch screen functionality (Distinguished project; 2015 HKUST President's Cup Winner!) |
| 2015 PAN4 Presentation Tools with Gesture Recognition and Augmented Reality |
Cool and dynamic presentation software using a Leap Motion controller and adding Augmented Reality (AR) into the presentations; includes an add-in for MS PowerPoint (Distinguished project) |
| 2015 PSAN1 AirTennis: A Web-Based, Mobile Motion Controlled Console Game |
A cool game that allows two players to play virtual tennis via computer displays and "tennis raquets" that are actually smart phones with accelerometers and gyroscopes (Distinguished project) |
| 2020 RO3 New Trading Strategies Inspired from Harmonic and Elliott Wave Trading Strategies |
The Harmonic Trading Strategy and related Elliott Wave Theory have led to revolutions in the field of Forex trading. The team implemented both strategies on the Quantopian platform and found that neither strategy is remotely as powerful when applied to other asset classes, particularly equities. They fall victim to the highly volatile market noise, often ending up losing more money than hands-off passive strategies. Thus, they tried a new strategy that incorporates both Harmonic and Elliott patterns, but is specialized for trading in the equities market, as it can adapt to uncertain market fluctuations. | Final Report |
| 2016 SC4 Electronic Health Record System for Developing Countries |
This system facilitates normal patient record-keeping work for NGOs and volunteer medical personnel who periodically visit rural areas with no Internet conectivity. Thanks to a Rasberry Pi single board computer, a LAN can be quickly set up for doctors and medical staff to immediately retrieve, modify and add medical records of patients in rural areas. The system is a lot more compact, portable, reliable, easy to understand and easy to use than previous solutions. The Android and iOS mobile apps eliminate the need of medical teams to carry notebook computers and power supplies to remote areas. The system also includes a dedicated API for better security and future scaling up. The team worked with SIGHT from HKUST and One2One CAMBODIA, and one team member actually spent several weeks in Cambodia gathering user requirements and field testing the system. Much documentation was also provided for the benefit of real-life end users. (Distinguished project) | Final Report |
| 2015 YIKE2 Flooding Simulation with Improved Tools and Features |
Flood simulation for a real terrain over a period of time under rainfall and other causes of flooding; utilizes the Shallow Water Equation to predict the water flow and return the water heights at respective locations over a certain time frame |
| 2020 KWT3 Wireless VR Game with Gesture and Voice Recognition |
This VR game integrates with a wireless VR headset, IBM Watson, Microsoft Speech Platform, Leap Motion, an accelerometer and fan to free the player from the restraints of traditional devices and allow more "reality" to be put into the game. |
| 2020 LZ1 Liveness Detection in Facial Recognition for a Security Gateway |
In order to detect whether people's skin is real or not, spectroscopic cameras with infrared light can be used to analyze the liveness of their faces. The basic theory of this project was that the reflection of lights on different materials differs significantly. Both RGB and NIR (near-infrared) algorithms were applied in this project. An RGB algorithm was used for detecting flesh color, while an NIR algorithm was used for detecting human skin. With the combination of both algorithms, only real human faces are detected. Three spectroscopic cameras with different filters were used to capture NIR light, while a normal digital camera was used to capture visible light. Lighting equipment that emits different wavelengths of lights was also applied to enhance the reflection. Hyperspectral data, as well as three-channel RGB image data, was collected and processed to reach the results. The project was sponsored and supported by Yitu Technology. |
| 2020 LZ2 Smart Traffic Signals Using Deep Learning-based Computer Vision |
This project attempted to utilize deep learning-based computer vision to develop smart traffic signals that can adapt to the current traffic conditions and optimize green duration for each lane in an intersection, resulting in a reduction of the average waiting time of vehicles and improvements in the traffic efficiency. |
| 2020 QUAN1 DepthFusion - A Study and Implementation of Lightweight Real-time 3D Reconstruction Algorithms Using Time-of-flight Depth Maps |
This project involved the study of current depth-map-based reconstruction algorithms. The result was a revised and improved program that can take in two types of inputs — 1) A series of depth maps and the corresponding translations and rotations (represented in quaternions) of the depth camera in the six degrees of freedom (6DoF), 2) a series of depth maps only — that can run solely with the CPU power of a laptop. The program was further improved by revising the code's time complexity, parameters and data structure based on past publications. The generated point clouds of the two outputs were compared and evaluated from 5 datasets to ensure the correctness of the pose estimation. The design of the algorithm was evaluated to provide an extensive study on how and where the GPU could be potentially used for parallelism to increase the efficiency of computation. |
| 2020 QUAN2 Avatar Face Reconstruction via RGB-D |
In this project, the team made a face reconstruction iOS application with the use of iPhone X's Truth-Depth camera. They built a light and simple program which significantly simplifies the face reconstruction process and is easily accessible, while keeping certain level of output quality. It can be used by 3D artists to replace part of the complex modeling procedure. The result of this project is a low-cost tool for making 3D avatars. |
| 2020 TA1 Wonder Painter: A General-purpose 3D Animation Platform |
In this project, the team developed a sketch-based 2D to 3D modeling and animating application. It was inspired by Teddy, a sketching interface for 3D freeform design, and by As-rigid-as-possible shape manipulation. The system contains three stages: Painting, Modeling, and Animating. Simple interfaces are provided in each stage for users to create 3D animation from scratch at ease. The system utilizes OpenFrameworks, an open-source C++ toolkit designed to wrap together several commonly used libraries (including OpenGL). The platform is user-friendly and enables normal people to make general-purpose 3D animations. |
| 2020 TM1 Automated Vision-based Wellness Analysis of Elderly Care Center Citizens |
This system provides a means to measure the wellness of people in an elderly care center. It utilizes cutting-edge deep learning models to automatically analyze spatio-temporal data, and it has two general components: facial analysis and activity analysis. FAcial analysis looks closely at facial information, identifies and tracks facial actions that occur within a time period and explores patterns that form through visualizations of the results. Activity analysis estimates the physical movements of elderly care center citizens during physical exercises, calculates the physical intensity of these citizens during the activities and visualizes as well as analyzes the patterns which form from the results. Two insights gained were that sleeping too much and blinking significantly more often may be signs of mild cognitive impairments. (Distinguished project) |
| 2020 TM2 Food Waste Analytics |
This system performs automatic monitoring of the food waste condition in Maxim’s Catering at LG1 of HKUST. The system detects the returned trays from videos captured at the tray return areas of the canteen. Then it classifies the dishes on each tray and estimates the amount of food leftovers on each dish. Visualizations are designed to present the findings in a more intuitive manner. To train the food classification model, 1610 tray images were labelled, which included 2292 food instances. The system achieved high accuracy in tray detection, dish classification and general food waste estimation. It was designed with portability and maintainability in mind, so applying it at another canteen environment should be very straightforward. The whole system is modularized to allow users to integrate their own machine learning models into our system, and it is dockerized to improve the stability. |
| 2020 TM4 Food Waste Analytics |
Based on video data collected from LG1 canteen in HKUST, this analytical web application analyzes food waste. It incorporates various computer vision techniques, such as object detection and semantic segmentation. This goal is to provide some meaningful information for vendors to customize their menus in an attempt to scale back the amount of leftovers. |
| 2020 TM5 Food Waste Segmentation and Volume Estimation |
In order to help reduce food waste at the LG1 canteen at HKUST, this system analyzes food and food waste on returned trays. With the help of object detection, text recognition and image segmentation techniques, the system discovers customers' eating habits and food preferences. These discoveries and statistics can help the restaurant management to formulate food serving strategies that reduce food waste to a minimum. |
| 2015 YIKE2 Flooding Simulation with Improved Tools and Features |
Flood simulation for a real terrain over a period of time under rainfall and other causes of flooding; utilizes the Shallow Water Equation to predict the water flow and return the water heights at respective locations over a certain time frame |
Copyright HKUST CSE Dept. 2024
Blog template built for Bootstrap
by @mdo.
Back to top