Current Projects
- 3D Gait Recognition
- Adaptive Resource Allocation Strategies for Emergency Management
- Analysis of Human Activities from 3D Data
- Deep Independently Recurrent Neural Networks (IndRNN)
- Early Detection and Assessment of Smoke from Single Gray-Scale Images
- Efficiently Learning Large-sized Symmetric Positive Definite Visual Representation
- Learning kernel-based high-order visual representation for image retrieval
- Local Descriptor based Image-to-Class Measure for Few-shot Learning
- Multi-Agent Social Learning for Market Prediction
- A Multi-Agent Solution for Smart Grid System Modelling, Simulation, and Management by Considering Distributed and Renewable Energy Resources and Local Power Storages
- Real-time Postural Assessment from Images
- A Self-boosting Framework for Automated Radiographic Report Generation
About
This project developed a technology for 3D gait recognition by using low-cost commodity 3D sensors. Evaluation of the method on a real-dataset of 266 subjects has shown that the method achieved 90% accuracy when using only a few training samples per subject The technology can be used as a person recognition, identification and re-identification. It is available for commercialisation or further extension.
Contact
About
In emergency management, one of the most important phase is to allocate human resource or materials to mitigate the damage caused by the events. For small-scale emergency events, public emergency services like police force, are department and ambulances can be used to control the damage. But for larger-scale emergency events, the damage control is obviously beyond the capability of these public emergency services. For examples, after a strong earthquake that happened in highly inhabited areas, thousands of people could be killed or affected. Under such situation, the search and rescue in the first few days is extremely important for victims, and it is obvious that separated public emergency services or departments cannot make efficient response if there is no applicable strategies to allocate or coordinate them. One promising way to make fast and effective resource allocation for large-scale emergency events is to use computer science and information technology.
Contact
About
Analysis of human motion such as recognition of actions and activities has a wide range of applications in video analytic for surveillance, robotics, health monitoring, diagnosis of neurological disorders, sports coaching and synthesis of human motion in VR/AR. By utilizing the fast-advanced low-cost 3D sensors, this project focuses on developing novel approaches to analysing 3D human motion from coarse to fine and micro levels. A number of awarding winning deep neural networks based technologies have been developed for action classification, semantic and zero-shot recognition and unsupervised recognition. Over 45 technical papers have been published in the top tier conferences and journals to date.
Contact
About
Independently Recurrent Neural Networks (IndRNN) is a type of Recurrent Neural Networks, which has a specific mechanism to overcome the issues of gradient exploding and vanishing problems. It is much more effective than LSTM and GRU. IndRNN also reduces the computation at each time step and can be over 10 times faster than the Long short-term memory (LSTM). Experiments have shown that IndRNN is able to process long sequences (over 5000 steps) and construct deep networks (over 100 layers). Better performance has been achieved on various tasks compared with the traditional RNN and LSTM. IndRNN is now widely used for prediction in many applications including agriculture, energy, health, transportation and finance.
Resources
Code is available at https://github.com/Sunnydreamrain/IndRNN_pytorch.
Contact
About
A novel and effective method was developed for detecting and separating smoke from a single image frame based on the atmospheric scattering models. The method separates a frame into quasi-smoke and quasi-background components and constructs novel features for early detection, localization and assessment of the severity of smoke. Empirical validation has shown that the method significantly outperforms existing features for early smoke detection. In particular, the proposed method is able to differentiate to a large extend smoke from other challenging objects (e.g. fog/haze, cloud, etc.) with similar visual appearance in a gray-scale frame. The technology is applicable for images taken by Drones, UVA and Satellite and is available for commercialisation.
Contact
About
Symmetric positive definite (SPD) matrix has recently been used as an effective visual representation. However, learning this representation in deep networks could result in significant computational cost. This project proposes a novel scheme called Relation Dropout (ReDro). Instead of using a full covariance matrix as in the literature, we generate a block diagonal one by randomly grouping the channels and only considering the covariance within the same group. As experimentally demonstrated, for the SPD methods typically involving the matrix normalisation step, ReDro can effectively help them reduce computational cost in learning large-sized SPD visual representation and also help to improve image recognition performance.
Contact
About
Image retrieval plays a key role in many practical applications. The recent increase of real-world applications calls for higher retrieval accuracy. This project aims to address this issue by exploring advanced visual representation that models the high-order information of image content. This project expects to generate new knowledge in the area of computer vision by developing a novel image retrieval framework. Expected outcomes include theory development on visual representation and more effective retrieval techniques. This should provide significant benefits, such as improving public information access services, facilitating environmental monitoring, and enhancing smart traffic management.
Contact
About
Few-shot learning in image classification aims to learn a classifier to classify images when only few training examples are available for each class. This project proposes a Deep Nearest Neighbour Neural Network (DN4 in short) and train it in an end-to-end manner. The proposed DN4 not only learns the optimal deep local descriptors for the image-to-class measure, but also utilises the higher efficiency of such a measure in the case of example scarcity, thanks to the exchangeability of visual patterns across the images in the same class. Our work leads to a simple, effective, and computationally efficient framework for few-shot learning.
Contact
About
In a large complex Smart Grid market, usually brokers are introduced for power trading. We design an intelligent broker agent—GongBroker, basing on the platform of Power TAC, which simulates a real world Smart Grid market. GongBroker buys energy from the wholesale market through auctions, and sells energy to various consumers in retail market. In our design, we deeply explore the customer consumption behaviors with a data-driven method. All the consumers are clustered according to their historical energy consumptions, and one-day-ahead hourly usage of each consumer cluster is predicted by Logistic regression. Basing on the predicted usages, the auction module employs an MDP process for one-day-ahead auction. To compete with other trading agents, GongBroker uses independent reinforcement learning processes to generate different tariffs for different types of consumers.
Contact
About
The objective of this project is to propose a multi-agent system to model, simulate and manage a Smart Grid (SG) system under different circumstances in automatic and efficient ways. Power management strategies will be proposed to dynamically balance the power supply and demand in the SG system, and self-healing mechanisms will be generated to intelligently diagnose the faults and to automatically restore the system from outages. The significance of this project lies in its promise to develop a new infrastructure for intelligent modelling, simulating and managing a SG system by considering reliability, efficiency, environmental and economy issues.
Contact
About
Humans face the risk of injuries caused by poor posture where body parts are at awkward or extreme positions which increase stress either in their daily activities and workplaces. Manual postural assessment by an ergonomist based on biomechanical principles is the common approach to early identification of the risks. This project focuses on postural assessment from images, aiming at a fully automatic, low cost, convenient, and reliable method/system for use in workplaces and daily lives. We have developed deep learning methods for upper-body and whole-body postural assessment. Compared to the manual assessment using “Rapid Entire Body Assessment” (REBA) criteria, our method achieved average Kappa Statistics 0.742 for the entire body and Kappa Statistics for individual body parts (Legs: 0.816, Lower arms: 0.788, Neck: 0.643, Trunk: 0.686, Upper arms: 0.815, Wrists: 0.702).
Contact
About
Automated radiographic report generation is a challenging task since it requires to generate paragraphs describing fine-grained visual differences of cases, especially for those between the diseased and the healthy. This project proposes a self-boosting framework that improves radiographic report generation based on the cooperation of the main task of report generation and an auxiliary task of image-text matching. The two tasks are built as the two branches of a network model and influence each other in a cooperative way. These two branches are jointly trained to help improve each other iteratively and progressively, so that the whole model is self-boosted without requiring external resources.