A Review of Skeleton-Based Human Action Recognition

Video-based human action recognition is one of the most important and challenging areas of research in the field of computer vision. Human action recognition has found many pragmatic applications in video surveillance, human-computer interaction, entertainment, autonomous driving, etc. Owing to the recent development of deep learning methods for human action recognition, the performance of action recognition has significantly enhanced for challenging datasets. Deep learning techniques are mainly used for recognizing actions in images and videos comprising of Euclidean data. A recent development in deep learning methods is the extension of these techniques to non-Euclidean data or graph data with many nodes and edges. Human body skeleton resembles a graph, therefore, the graph convolutional network (GCN) is applicable to the non-Euclidean body skeleton. In the past few years, GCN has emerged as an important tool for skeleton-based action recognition. Therefore, we conduct a survey using GCN methods for action recognition. Herein, we present a comprehensive overview of recent GCN techniques for action recognition, propose a taxonomy for the categorization of GCN techniques for action recognition, carry out a detailed study of the benchmark datasets, enlist relevant resources and open-source codes, and finally provide an outline for future research directions and trends. To the best of authors' knowledge, this is the first survey for action recognition using GCN techniques.

This paper presents a novel framework for skeleton-based Human Action Recognition (HAR) based on Graph Convolution Networks (GCNs). The proposed framework aims to increase human action recognition performance of GCN-based methods by incorporating a missing-joint-handling pre-processing step and a novel adjacency matrix construction method in a single human action recognition pipeline. The missing-joint-handling pre-processing step is utilized to infer missing data in the input sequence, which may occur due to imperfect skeleton extraction, based on imputation methods. The novel adjacency matrix construction method is executed offline to compute an improved weighted adjacency matrix specifically designed for HAR, which is utilized in every layer of the employed GCN. Moreover, both the pre-processing step and the adjacency construction method can be utilized along with any GCN architecture, allowing any GCN-based HAR method to be employed in the proposed framework. Experimental evaluation on two public datasets indicate favorable human action classification scores compared to the employed baseline and all competing methods both for 2D and 3D skeleton-based human action recognition, while using a GCN architecture with less learnable parameters.

Skeleton-based human action recognition has great potential for human behavior analysis owing to its simplicity and robustness in varying environments. This paper presents a spatial and temporal attention-enhanced graph convolution network (STAEGCN) for human action recognition. The spatial-temporal attention module in the network uses convolution embedding for positional information and adopts multi-head self-attention mechanism to extract spatial and temporal attention separately from the input series of the skeleton. The spatial and temporal attention are then concatenated into an entire attention map according to a specific ratio. The proposed spatial and temporal attention module was integrated with an adaptive graph convolution network to form the backbone of STAEGCN. Based on STAEGCN, a two-stream skeleton-based human action recognition model was trained and evaluated. The model performed better on both NTU RGB+D and Kinetics 400 than 2s-AGCN and its variants. It was proven that the strategy of decoupling spatial and temporal attention and combining them in a flexible way helps improve the performance of graph convolution networks in skeleton-based human action recognition.

Human action abnormality detection has been attempted by various sensors for application domains like rehabilitation, healthcare, and assisted living. Since the release of motion sensors that ease the human body skeleton retrieval, skeleton-based human action recognition has recently been an active topic in the area of artificial intelligence. Unlike human action recognition, human action abnormality detection is an emerging field that aims to detect the incorrect action from the same action class. Graph convolutional network has been widely adopted for human action recognition. However, to the best of our knowledge, whether it could be effective for the task of human action abnormality detection has not been attempted. To advance prior work in the emerging field of human action abnormality detection, we propose a novel method that uses graph convolutional network to detect abnormal actions in skeleton data. To validate the effectiveness of our proposed method, we conduct extensive experiments on a public dataset called UI-PRMD. Based on the experimental results, our proposed method achieved superior action abnormality detection performance comparing with existing deep learning methods.

In recent years, graph convolutional networks (GCNs) play an increasingly critical role in skeleton-based human action recognition. However, most GCN-based methods still have two main limitations: 1) They only consider the motion information of the joints or process the joints and bones separately, which are unable to fully explore the latent functional correlation between joints and bones for action recognition. 2) Most of these works are performed in the supervised learning way, which heavily relies on massive labeled training data. To address these issues, we propose a semi-supervised skeleton-based action recognition method which has been rarely exploited before. We design a novel correlation-driven joint-bone fusion graph convolutional network (CD-JBF-GCN) as an encoder and use a pose prediction head as a decoder to achieve semi-supervised learning. Specifically, the correlation-driven joint-bone fusion graph convolution (CD-JBF-GC) can explore the motion transmission between the joint stream and the bone stream, so as to promote both streams to learn more discriminative feature representations. The pose prediction based auto-encoder in the self-supervised training fashion allows the network to learn motion representation from the unlabeled data, which is essential for action recognition. Extensive experiments on two popular datasets, i.e. NTU-RGB+D and Kinetics-Skeleton, demonstrate that our model achieves the state-of-the-art performance for semi-supervised skeleton-based action recognition and is also useful for fully-supervised methods.

Adaptive multi-view graph convolutional networks for skeleton-based action recognition

Healthcare using body sensor data has been getting huge research attentions by a wide range of researchers because of its good practical applications such as smart health care systems. For instance, smart wearable sensor-based behavior recognition system can observe elderly people in a smart eldercare environment to improve their lifestyle and can also help them by warning about forthcoming unprecedented events such as falls or other health risk, to prolong their independent life. Although there are many ways of using distinguished sensors to observe behavior of people, wearable sensors mostly provide reliable data in this regard to monitor the individual’s functionality and lifestyle. In this paper, we propose a body sensor-based activity modeling and recognition system using time-sequential information-based deep Neural Structured Learning (NSL), a promising deep learning algorithm. First, we obtain data from multiple wearable sensors while the subjects conduct several daily activities. Once the data is collected, the time-sequential information then go through some statistical feature processing. Furthermore, kernel-based discriminant analysis (KDA) is applied to see the better clustering of the features from different activity classes by minimizing inner-class scatterings while maximizing inter-class scatterings of the samples. The robust time-sequential features are then applied with Neural Structured Learning (NSL) based on Long Short-Term Memory (LSTM), for activity modeling. The proposed approach achieved around 99% recall rate on a public dataset. It is also compared to existing different conventional machine learning methods such as typical Deep Belief Network (DBN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) where they yielded the maximum recall rate of 94%. Furthermore, a fast and efficient explainable Artificial Intelligence (XAI) algorithm, Local Interpretable Model-Agnostic Explanations (LIME) is used to explain and check the machine learning decisions. The robust activity recognition system can be adopted for understanding peoples' behavior in their daily life in different environments such as homes, clinics, and offices.

Skeleton-based action recognition aims to project skeleton sequences to action categories, where skeleton sequences are derived from multiple forms of pre-detected points. Compared with earlier methods that focus on exploring single-form skeletons via Graph Convolutional Networks (GCNs), existing methods tend to improve GCNs by leveraging multi-form skeletons due to their complementary cues. However, these methods (either adapting structure of GCNs or model ensemble) require the co-existence of all forms of skeletons during both training and inference stages, while a typical situation in real life is the existence of only partial forms for inference. To tackle this issue, we present Adaptive Cross-Form Learning (ACFL), which empowers well-designed GCNs to generate complementary representation from single-form skeletons without changing model capacity. Specifically, each GCN model in ACFL not only learns action representation from the single-form skeletons, but also adaptively mimics useful representations derived from other forms of skeletons. In this way, each GCN can learn how to strengthen what has been learned, thus exploiting model potential and facilitating action recognition as well. Extensive experiments conducted on three challenging benchmarks, i.e., NTU-RGB+D 120, NTU-RGB+D 60 and UAV-Human, demonstrate the effectiveness and generalizability of the proposed method. Specifically, the ACFL significantly improves various GCN models (i.e., CTR-GCN, MS-G3D, and Shift-GCN), achieving a new record for skeleton-based action recognition.

In recent years, human action recognition has received increasing attention as a significant function of human–machine interaction. The human skeleton is one of the most effective representations of human actions because it is highly compact and informative. Many recent skeleton-based action recognition methods are based on graph convolutional networks (GCNs) as they preserve the topology of the human skeleton while extracting features. Although many of these methods give impressive results, there are some limitations in robustness, interoperability, and scalability. Furthermore, most of these methods ignore the underlying information of view direction and rely on the model to learn how to adjust the view from training data. In this work, we propose VW-SC3D, a spatial–temporal model with view weighting for skeleton-based action recognition. In brief, our model uses a sparse 3D CNN to extract spatial features for each frame and uses a transformer encoder to obtain temporal information within the frames. Compared to GCN-based methods, our method performs better in extracting spatial–temporal features and is more adaptive to different types of 3D skeleton data. The sparse 3D CNN makes our model more computationally efficient and more flexible. In addition, a learnable view weighting module enhances the robustness of the proposed model against viewpoint changes. A test on two different types of datasets shows a competitive result with SOTA methods, and the performance is even better in view-changing situations.

In recent years, many researchers have focused on using convolutional neural networks to perform human activity recognition as evidenced by the emergence of a number of convolutional neural network architectures such as LeNet-5, AlexNet and VGG16 and modern architectures such as ResNet, Inception V3, Inception-ResNet, MobileNet V2, NASNet and PNASNet. The main characteristic of a convolutional neural network (CNN) is its ability to extract features automatically from input images, which facilitates the processes of activity recognition and classification. Convolutional networks indeed derive more relevant and complex features with every additional layer. In addition, CNNs have achieved perfect classification on highly similar activities that were previously extremely difficult to classify. In this paper, we evaluate modern convolutional neural networks in terms of their human activity recognition accuracy, and we compare the results with the state of the art methods. In our research, we used two public data sets, HMDB (Shooting gun, kicking, falling to the floor, punching) and the Weizman dataset (walking, running, jumping, bending, one hand waving, two-hand waving, jumping in place, jumping jack, skipping). Our experimental results indicated that the CNN with NASNet architecture achieves the best performance of the six CNN architectures on both human activity data sets (HMDB and Weizman).

Skeleton-based human action recognition is a research hotspot in recent years, but most of the research focuses on the spatio-temporal feature extraction by convolutional neural network. In order to improve the correct recognition rate of these models, this paper proposes three strategies: using algebraic method to reduce redundant video frames, adding auxiliary edges into the joint adjacency graph to improve the skeleton graph structure, and adding some virtual classes to disperse the error recognition rate. Experimental results on NTU-RGB-D60, NTU-RGB-D120 and Kinetics Skeleton 400 databases show that the proposed strategy can effectively improve the accuracy of the original algorithm.

In recent years, Deep Learning (DL) models, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM), have been widely used for Human Activities Recognition (HAR). They have achieved considerable performance improvements over classical Machine Learning (ML) approaches due to their excellent feature representation capabilities. Simple human activities are performed in sequential order with no overlaps or concurrent actions. The paucity of labeled training activity samples, high computational cost, and system resource requirements of deep learning architectures compared to lightweight models are some of the research challenges confronting the HAR community. We framed the lightweight DL-based CNNs, RNNs, and LSTM model for HAR to tackle these research challenges. The RNNs and LSTM have single layers with softmax activation functions, whereas the one-dimensional CNNs have two convolutional and single max-pooling layers. To evaluate the performance of our models, we used the publicly available benchmark WISDM experimental dataset, which includes the reading of six activities (walking, jogging, upstairs, downstairs, sitting, and standing) performed by 36 participants using a single accelerometer sensor. The experimental result illustrates the efficacy of our models in activity recognition, demonstrating that they attain higher accuracy while being computationally efficient.

Graph convolutional networks (GCN), which can model the human body skeletons as spatial and temporal graphs, have shown remarkable potential in skeleton-based action recognition. However, in the existing GCN-based methods, graph-structured representation of the human skeleton makes it difficult to be fused with other modalities, especially in the early stages. This may limit their scalability and performance in action recognition tasks. In addition, the pose information, which naturally contains informative and discriminative clues for action recognition, is rarely explored together with skeleton data in existing methods. In this work, we proposed pose-guided GCN (PG-GCN), a multi-modal framework for high-performance human action recognition. In particular, a multi-stream network is constructed to simultaneously explore the robust features from both the pose and skeleton data, while a dynamic attention module is designed for early-stage feature fusion. The core idea of this module is to utilize a trainable graph to aggregate features from the skeleton stream with that of the pose stream, which leads to a network with more robust feature representation ability. Extensive experiments show that the proposed PG-GCN can achieve state-of-the-art performance on the NTU RGB+D 60 and NTU RGB+D 120 datasets.

In skeleton-based action recognition, graph convolutional networks (GCNs) have achieved remarkable success. Nevertheless, how to efficiently model the spatial-temporal skeleton graph without introducing extra computation burden is a challenging problem for industrial deployment. In this paper, we rethink the spatial aggregation in existing GCN-based skeleton action recognition methods and discover that they are limited by coupling aggregation mechanism. Inspired by the decoupling aggregation mechanism in CNNs, we propose decoupling GCN to boost the graph modeling ability with no extra computation, no extra latency, no extra GPU memory cost, and less than 10% extra parameters. Another prevalent problem of GCNs is over-fitting. Although dropout is a widely used regularization technique, it is not effective for GCNs, due to the fact that activation units are correlated between neighbor nodes. We propose DropGraph to discard features in correlated nodes, which is particularly effective on GCNs. Moreover, we introduce an attention-guided drop mechanism to enhance the regularization effect. All our contributions introduce zero extra computation burden at deployment. We conduct experiments on three datasets (NTU-RGBD, NTU-RGBD-120, and Northwestern-UCLA) and exceed the state-of-the-art performance with less computation cost.

Human action recognition is one of the raison d’être for doing human–computer interaction research, as it is highly vital in meeting the demands of modern society, such as automatic video surveillance for security, patient monitoring for recovery, content-based video retrieval, etc. In line with this, deep learning systems are fast becoming the defacto standard for object recognition, video understanding, and pattern recognition due to their inherent powerful feature learning ability from vast amount of data. It makes sense to capitalize on its great success and to further improve it for the complex task of action recognition. One of the contributions in this paper is an effective and yet simple method for encoding the spatiotemporal information from skeleton sequences into what we call temporal kinematic images. In the input encoding scheme, we embed various geometric relational features derived from the skeleton sequence in the form of our proposed skeletal optical flows (SOFs). SOFs collectively represent the variations of kinetic energy, angles between limbs, and pair-wise displacements between joints over consecutive frames of skeleton data, as color variations in the temporal kinematic images. Another contribution is our convolutional neural network with a correctness-vigilant regularizer. It is employed to exploit the discriminative features from the temporal kinematic image for human action recognition. Lastly, we additionally investigated an adaptive label smoothing technique employed toward the end of training iterations. Empirical results show that the efficiency of the proposed method is superior to existing works in terms of the generalizability of the generated model, training convergence speed, and the resulting classification accuracy on nine popular benchmarking datasets, such as MHAD, MSR Activity 3D, HDM05, MSR Daily Activity 3D, and the latest challenging databases, such as UTKinect-Action, NTU RGB+D, Northwestern-UCLA, UWA3DII, and SBU Kinect Interaction datasets.