25 March 2021 (1.00pm, Thursday, Online) Dr Jia Lu (CS, AUT) Topics on “Deep Learning Methods for Human Behaviour Recognition”
Video surveillance has been broadly applied to public places which allows security staff to monitor abnormal events. However, most of the video surveillance software platforms are still being run in traditional mode. As the state-of-the-art technology, deep learning (DL) becomes much popular because of its superiority to conventional machine learning. Moreover, deep learning as an end-to-end model normally does not require low-level processing which is able to cut off human labour and gain time efficiency for human behaviour recognition though the training is costly. In this project, we design and implement deep learning methods which are time efficiency and outperform in training and testing. As the outcome of this research work, we attain an overall up to 90% accuracy of recognition in real-time. Moreover, we have created our own dataset for the experiments.
23 March 2021 (1.00pm, Tuesday, Online) Ms Xuejing Chen (CS, AUT) Topics on “Deep Learning Approach to Spatio-Temporal Image Recognition for Road Surface Damage Detection”
Road damages like cracks, alligators, and potholes are critical problems for road constructions and major considerations for driving decisions. Road pavement distresses, at the early stage of road damage formation, also signal needs for road maintenance. Recently computer-vision techniques for detecting such damages mostly rely on convolutional neural networks which have been studied but not reached good results, yet still facing the limitations related to localisation, 3D consideration, and run-time efficiency. In this research project, we propose a deep learning architecture to efficiently learn, detect, and reconstruct 3D road damages concerning different quality and temporal-spatial information of streaming images. We have conducted an experimental study and the results show that our initially proposed architecture of Ensembled Aggregated Pyramid Residual Network achieves better performance than the state-of-the-art CNN-based models. Based on this, we continue improving and evaluating our models based on BDD and CCSAD to extract temporal spatial information as well as 3D images of road damage objects. We will also perform temporal-spatial causality analysis and stereo measures to obtain a real-time and fine-grained road surface damage measurement. Our research work will advance the field and can help transportation agencies make informed decisions on scheduling a repair or rebuilding at an appropriate time.
11 Feb 2021 -22 April 2021 (12.00pm, Thursday, Online) Wei Qi Yan (CS, AUT) Topics on “Postgraduate Mathematics and Reinforcement Learning”
(1) Functional Analysis: Part 1 (11 February 2021)
(2) Functional Analysis: Part 2 (18 February 2021)
(3) Finite Fields (25 February 2021)
(4) Manifold Learning (04 March 2021)
(5) Probabilistic Graphical Models (11 March 2021)
(6) Generative Adversarial Networks (18 March 2021)
(7) Autoencoder (25 March 2021)
(8) Reinforcement Learning (01 April 2021)
(9) RL1-1: Exact Solution Methods (08 April 2021)
(10) RL1-2: Exact Solution Methods (15 April 2021)
(11) RL2: Approximate Solution Methods (22 April 2021)