Developing Advanced Deep Learning Algorithms for Video-based Human Action Recognition

人类动作和行为识别对于工业和医疗保健应用至关重要。静态图像的分析已经取得了一定的进展，然而，许多行为不能被识别的静态图像，因为无法捕捉的行为运动的时间信息。随着科学技术的进步和人们生活的便利，视频文件变得容易获取。视频文件记录了人类行为和动作的时间顺序，发生的对象以及它们发生的环境。最近，基于深度学习的方法在视频动作识别任务中取得了巨大的成功。然而，在实际应用中也存在着一些相当大的问题，一方面，视频视点和视频外观的多样性，导致实际应用中的准确性大大降低。另一方面，视频处理涉及3D网络来处理图像帧，这将产生大量参数和计算开销。反过来，这在速度和存储器方面给计算硬件带来了问题，并且过大的网络不适合实际应用。因此，该项目将侧重于开发机器学习和高性能计算方法，以准确有效地识别人类行动和行为。深度学习技术将被研究用于更小的网络架构的高识别性能。将研究新的计算方法，以通过GPU加速这一过程。这些模型将由公共和私人数据集开发和验证。

Human action and behaviours recognition is essential for industries and healthcare applications. Some progress has been made on the analysis of static images, however, many behaviours cannot be recognised by still images because of the inability of capturing the temporal information of the behaviour motion. With the advances of science and technology and the convenience of people's life, video files have become easy to obtain. Video files record the time sequence of human behaviours and actions, the objects of occurrence, and the environment in which they occur. Recently, the deep learning-based method show great success in the video action recognition task. However, there are some considerable problems in applications, On the one hand, the variety of video viewpoints and video appearance, lead to a significant decrease in the accuracy in practical application. On the other hand, video processing involves a 3D network to process frames of images, which will generate a lot of parameters and computational overhead. In turn, this poses a problem to computing hardware in terms of speed and memory and an excessively large network is not suitable for practical application. Thus, this project will focus on the development of machine learning and high-performance computing methods for the accurate and effective recognition of human action and behaviour. Deep learning techniques will be investigated towards high recognition performance with smaller network architecture. New computing approaches will be studied to speed up the process via GPU. The models will be developed and validated by public and private datasets.