动作识别是计算机视觉研究中的一个基本但具有挑战性的问题。在过去的几年中，许多基于RGB视频的识别技术已经得到了巨大的发展，并取得了显著的成果。但是，处理RGB视频可能非常耗时。其中，在动作识别领域，人体骨架数据具有轻量级的特点，同时对人体外观、环境背景等信息具有不变性，因此，这种数据模态受到了越来越多的关注。然而，基于人体骨架的动作识别面临两个问题：人体骨架数据的噪声问题和数据标注的依赖问题。噪声问题是指骨架数据中存在噪声影响数据的准确性，而数据标注依赖问题则是指在监督学习中，需要大量的标签数据进行训练。本文针对人体骨架数据在采集中的噪声问题，提出了一种基于噪声适应的动作识别模型，设计了回归模型和生成模型充分利用不同场景下的噪声数据特点。并且针对人体骨架数据过于依赖标签数据，利用自监督学习方法，提出了一个基于多任务自监督学习的动作识别方法。

Action recognition is a fundamental yet challenging problem in computer vision. In the past few years, many works have been developed on recognition based on RGB videos and achieved many significant results. However, processing RGB videos can be very time consuming. Another data modality, human skeletons, which represent a person by the 3D coordinate positions of skeletal joints, draw much attention due to the lightweight representations, the robustness to variations of viewpoints, appearances, and surrounding distractions. However, action recognition of skeleton data faces two problems: noise of skeleton data and dependence of data annotation. The problem of noise refers to the noise in skeleton data that affects the accuracy of data, while the problem of data annotation dependence refers to that the training requires lots of labelled data. To address the issue of action analytics from noisy skeletons which commonly appear in the real world, this paper proposes a noise-adaptation model to get rid of explicit skeleton noise modelling and reliance on skeleton ground truths. Regression-based and generation-based adaptation models are developed respectively according to whether the pairs of noisy skeletons are available. Besides, aiming at dependence of data annotation, with the self-supervised learning method on human skeleton data, this paper proposes an action recognition method based on multitask self-supervised learning.