基于双流网络的人体行为特征识别研究

A Review of Human Action Feature Recognition Based on Two-Stream Convolutional Networks

摘要:

人体行为特征识别作为计算机视觉领域的一个重要研究方向，在实际生活中有着广泛的应用，其研究方法可以分为基于传统方法和基于深度学习的方法；双流网络作为基于深度学习方法中较为经典的网络，其将视频序列分为时间和空间两种特征的思想为研究者们提供了丰富的思路；从基于3D-CNN的双流网络、融合LSTM的双流网络、基于图卷积的双流网络和引入注意力机制的双流网络四个方面分别探讨双流网络的研究现状，分析各类方法的局限性，梳理双流网络发展的关键节点并总结每种方法的优缺点及应用场景；列举常用数据集；概述人体行为特征识别的应用；指出目前面临的挑战并对未来进行展望。

Abstract:

Human action feature recognition, as an important research direction in the field of computer vision, has wide applications in real life. Its research methods can be divided into traditional methods and deep learning-based methods. The two-stream convolutional networks, as a classic network in deep learning-based methods, provides researchers with a rich set of ideas by dividing video sequences into two types of features: temporal and spatial. This paper investigates the current research status of two-stream convolutional networks for human action feature recognition from four aspects: two-stream convolutional networks based on 3D convolutional neural network, two-stream convolutional networks integrated with long short-term memory, two-stream convolutional networks based on graph convolutional network, and two-stream convolutional networks incorporating attention mechanisms. It analyzes the limitations of various methods, reviews the key milestones in the development of two-stream convolutional networks, and summarizes the advantages, disadvantages, and application scenarios of each method. The paper also lists commonly used datasets and provides an overview of the practical applications of human action feature recognition. Additionally, it identifies the challenges currently faced, and provides an outlook for the future.