Human Activity Recognition Dataset Research Articles

Low-light aware framework for human activity recognition via optimized dual stream parallel network

Human Activity Recognition (HAR) plays a crucial role in communication and the Internet of Things (IoT), by enabling vision sensors to understand and respond to human behavior more intelligently and efficiently. Existing deep learning models are complex to deal with the low illumination, diverse viewpoints, and cluttered backgrounds, which require substantial computing resources and are not appropriate for edge devices. Furthermore, without an effective video analysis technique it processes entire frames, resulting inadequate performance. To address these key challenges, a cloud-assisted IoT computing framework is proposed for HAR in uncertain low-lighting environments, which is mainly composed of two tiers: edge and cloud computing. Initially, a lightweight Convolutional Neural Network (CNN) model is developed which is responsible to enhance the low-light frames, followed by the human detection algorithm to process the selective frames, thus enabling efficient resource utilization. Next, these refined frames are then transmitted to the cloud for accurate HAR, where dual stream CNN and transformer fusion network extract both short-and long-range spatiotemporal discriminative features followed by proposed Optimized Parallel Sequential Temporal Network (OPSTN) with squeeze and excitation attention to efficiently learn HAR in complex scenarios. Finally, extensive experiments are conducted over three challenging HAR datasets to deeply examine the proposed framework from various perspectives such as complex activity recognition, lowlighting, etc., where the results are outperformed compared with the state-of-art methods.

Semisupervised Generative Adversarial Networks With Temporal Convolutions for Human Activity Recognition

Many potential applications of human activity recognition (HAR) can be found in health, surveillance, manufacturing, sports, etc. For instance, HAR can be exploited in ambient assisted living (AAL) systems to provide users with assistance services intend to improve their well-being, safety, and autonomy. Data annotation in HAR is a complex and time-consuming process that limits the availability of labeled samples. Furthermore, the classification performance of supervised deep neural networks depends on the availability of large, annotated training data. This paper proposes an alternative framework for semi-supervised generative adversarial networks (GANs) using temporal convolutions for semi-supervised action recognition to address several problems related to conventional approaches in the HAR context, such as high-dimensionality, scarcity of annotated data, scalability, and robustness. The proposed framework employs a single architecture on different datasets and its effectiveness is tested under four conditions reflecting real-world semi-supervised scenarios to investigate the impact of intersubject training, amount of labeled data, number of classes, and IMU positions on model performance. The evaluations are performed on the PAMAP2, Opportunity-locomotion, and LISSI HAR datasets to evaluate the generalizability of the framework. The results show the proposed framework’s high classification performance and generalization ability compared to baseline methods, achieving up to 25% improvement when only a small amount of annotated data is available. A comparison to the previous work using the same datasets has also validated the performance of the framework.

Counting Activities Using Weakly Labeled Raw Acceleration Data: A Variable-Length Sequence Approach with Deep Learning to Maintain Event Duration Flexibility.

This paper presents a novel approach for counting hand-performed activities using deep learning and inertial measurement units (IMUs). The particular challenge in this task is finding the correct window size for capturing activities with different durations. Traditionally, fixed window sizes have been used, which occasionally result in incorrectly represented activities. To address this limitation, we propose segmenting the time series data into variable-length sequences using ragged tensors to store and process the data. Additionally, our approach utilizes weakly labeled data to simplify the annotation process and reduce the time to prepare annotated data for machine learning algorithms. Thus, the model receives only partial information about the performed activity. Therefore, we propose an LSTM-based architecture, which takes into account both the ragged tensors and the weak labels. To the best of our knowledge, no prior studies attempted counting utilizing variable-size IMU acceleration data with relatively low computational requirements using the number of completed repetitions of hand-performed activities as a label. Hence, we present the data segmentation method we employed and the model architecture that we implemented to show the effectiveness of our approach. Our results are evaluated using the Skoda public dataset for Human activity recognition (HAR) and demonstrate a repetition error of ±1 even in the most challenging cases. The findings of this study have applications and can be beneficial for various fields, including healthcare, sports and fitness, human-computer interaction, robotics, and the manufacturing industry.

HARBIC: Human activity recognition using bi-stream convolutional neural network with dual joint time–frequency representation

Human activity recognition (HAR) based on wearable devices has progressively advanced in context-aware computing like healthcare, smart homes, and industry 4.0. Since machine learning and Internet-of-Things have been in the spotlight, many deep learning-based approaches have been proposed, achieving very good performance using inertial data. In this study, we explore different joint time–frequency representations of the sensor data, in conjunction with a Convolutional Neural Network (ConvNet) to implement activity classification. We also consider the concurrent use of two different data representations in a cooperative bi-stream ConvNet configuration. Scaling of joint time–frequency representation is explored to reduce the ConvNet model size, and hence the computational complexity of the proposed approach. The classification performance and energy efficiency of the proposed method are evaluated on a real-world public HAR dataset. The performance of pairwise combinations and also, the single-stream of the two prominent representations in the joint time–frequency domain are compared to that of the state-of-the-art. The proposed method is implemented on Google’s Edge TPU platform to investigate the trade-off between the energy footprint and classification performance, which is critical in IoT applications where machine learning inference needs to be performed on resource-constrained edge devices. The highest overall classification accuracy and F1 score achieved by our method was 98.39% and 96.86%, respectively, which represents a significant improvement over the state-of-the-art.

Enhancing Zero-Shot Action Recognition in Videos by Combining GANs with Text and Images

Zero-shot action recognition (ZSAR) tackles the problem of recognising actions that have not been seen by the model during the training phase. Various techniques have been used to achieve ZSAR in the field of human action recognition (HAR) in videos. Techniques based on generative adversarial networks (GANs) are the most promising in terms of performance. GANs are trained to generate representations of unseen videos conditioned on information related to the unseen classes, such as class label embeddings. In this paper, we present an approach based on combining information from two different GANs, both of which generate a visual representation of unseen classes. Our dual-GAN approach leverages two separate knowledge sources related to the unseen classes: class-label texts and images related to the class label obtained from Google Images. The generated visual embeddings of the unseen classes by the two GANs are merged and used to train a classifier in a supervised-learning fashion for ZSAR classification. Our methodology is based on the idea that using more and richer knowledge sources to generate unseen classes representations will lead to higher downstream accuracy when classifying unseen classes. The experimental results show that our dual-GAN approach outperforms state-of-the-art methods on the two benchmark HAR datasets: HMDB51 and UCF101. Additionally, we present a comprehensive discussion and analysis of the experimental results for both datasets to understand the nuances of each approach at a class level. Finally, we examine the impact of the number of visual embeddings generated by the two GANs on the accuracy of the models.

ICA and IVA bounded multivariate generalized Gaussian mixture based hidden Markov models

Machine learning (ML), a branch of artificial intelligence (AI), is an area of computational science that is concerned with the analysis and interpretation of patterns and structures in data to enable learning and decision-making without the participation of a human. Hidden Markov models (HMMs), which have been acknowledged for decades but have recently made a significant revival in machine learning, are one of the most impressively powerful probabilistic models. HMMs are frequently employed in machine learning to model heterogeneous time series data. In this paper, we integrate independent component analysis (ICA) and ICA with a bounded multivariate generalized Gaussian mixture model (ICA-BMGGMM) into the HMM approach. One limitation of ICA is that it assumes the sources to be independent from each other. This assumption can be relaxed by combining independent vectors analysis (IVA) and IVA with the BMGGMM (IVA-BMGGMM) into the HMM approach to improve their modeling capability. We validate our proposed models using a variety of applications, such as human action recognition, speech recognition, and energy disaggregation. The results presented in the paper demonstrate the effectiveness of the proposed approaches for modeling different types of data. These data include KTH and Weizmann datasets for human action recognition, TIMIT and SDR for speech recognition, REDD dataset for energy disaggregation and EEG dataset for elliptic seizure classification. For all conducted experiments, our proposed models outperform other comparing models for all performance metrics such as accuracy, sensitivity, and precision. The best detection results were found using the IVABMGGMM-HMM for the reported experiments.

Transformer-based deep reverse attention network for multi-sensory human activity recognition

In today’s era, one of the important applications of Artificial Intelligence (AI) is Human Activity Recognition (HAR). It has a wide range of applicability in health monitoring for patients with chronic diseases, gaming consoles for gesture recognition, etc. Sensor-based HAR systems use signals collected over a period of time to label an activity. When we design an efficient sensor-based HAR system, a model requires learning an optimal association of spatial and temporal features. In this article, we propose a sensor-based HAR technique using the deep learning approach. We present a deep reverse transformer-based attention mechanism to guide the side residual features Unlike the conventional bottom-up approaches for feature fusion, we exploit a top-down feature fusion approach. The reverse attention is self-calibrated throughout the course of learning, which regularizes the attention modules and dynamically adjusts the learning rate. The overall framework outperforms several state-of-the-art methods and is shown to be statistically significant against these methods on five publicly available sensor-based HAR datasets, namely, MHEALTH, USC-HAD, WHARF, UTD-MHAD1, and UTD-MHAD2. Further, we conduct an ablation study to showcase the importance of each of the components of the proposed framework. Source code of this work is available at https://github.com/rishavpramanik/RevTransformerAttentionHAR.

Probabilistic knowledge infusion through symbolic features for context-aware activity recognition

In the general machine learning domain, solutions based on the integration of deep learning models with knowledge-based approaches are emerging. Indeed, such hybrid systems have the advantage of improving the recognition rate and the model’s interpretability. At the same time, they require a significantly reduced amount of labeled data to reliably train the model. However, these techniques have been poorly explored in the sensor-based Human Activity Recognition (HAR) domain. The common-sense knowledge about activity execution can potentially improve purely data-driven approaches. While a few knowledge infusion approaches have been proposed for HAR, they rely on rigid logic formalisms that do not take into account uncertainty. In this paper, we propose P-NIMBUS, a novel knowledge infusion approach for sensor-based HAR that relies on probabilistic reasoning. A probabilistic ontology is in charge of computing symbolic features that are combined with the features automatically extracted by a CNN model from raw sensor data and high-level context data. In particular, the symbolic features encode probabilistic common-sense knowledge about the activities consistent with the user’s surrounding context. These features are infused within the model before the classification layer. We experimentally evaluated P-NIMBUS on a HAR dataset of mobile devices sensor data that includes 14 different activities performed by 25 users. Our results show that P-NIMBUS outperforms state-of-the-art neuro-symbolic approaches, with the advantage of requiring a limited amount of training data to reach satisfying recognition rates (i.e., more than 80% of F1-score with only 20% of labeled data).

TALK: Tracking Activities by Linking Knowledge

Dependable and accurate monitoring of elderly at home becomes crucial to limit both the costs and human efforts of following up elderly for establishing a healthy care system. Human Activity Recognition (HAR) tools, based on sensors installed in smart homes, will become an important tool to provide useful information to the caregiver when something happens in the house of an elderly and care is required. The current available detection tools either exist out of interpretable knowledge-driven techniques or scalable data-driven ones. In this paper, a hybrid methodology that combines both approaches is designed and evaluated to Track Activities by Linking Knowledge (TALK). Both sensor data and their link to the relevant domain knowledge about where those sensors are installed, the performed activities that occur, and how the household is constructed, are generalized in a specific knowledge graph (KG) structure to represent continuous events. The interpretable knowledge graph embedding technique Instance Neighboring using Knowledge (INK) is then used to transform these events inside the KG to a tabular format, which can be used by any traditional machine learning classifier to create a HAR tool. The TALK methodology is evaluated on two HAR datasets and shows (a) that TALK outperforms both traditional automated data-driven as well as knowledge-driven techniques in terms of predictive performance, and (b) how TALK can be easily used in a more out of lab environment. All these results and the interpretable aspects show that TALK can become an important tool to monitor elderly in their homes efficiently, effectively and with less intrusive techniques.

Aeriform in-action: A novel dataset for human action recognition in aerial videos

Human actions being diverse in nature cannot be generalized, thus making it quite difficult to train a machine to recognize such diversified actions. This challenge is further compounded by the lack of availability of datasets for aerial surveillance, as collecting and annotating a large dataset is a formidable task. This paper aims to solve the problem of data scarcity by introducing a new dataset, Aeriform in-action for recognizing human actions from aerial videos. The proposed dataset consists of 32 high-resolution videos containing 13 action classes with 55,477 frames (without augmentation) and almost 400,000 annotations. It includes complex and aggressive actions such as kicking and punching, as well as drone signaling actions like waving and handshaking. The dataset also includes human-object interactions like carrying and reading. In addition to the dataset, this paper also presents a two-step deep learning framework for recognizing human actions based on the integration of human detection and action recognition module. The action recognition module adopts a modified version of the ResNeXt101 architecture (M-ResNext101) to recognize human actions in aerial videos. The performance of the proposed M-ResNext101 model is compared with 13 other deep learning models, and it outperforms all of them with an accuracy of 76.44% on the test data. The proposed dataset for human action recognition in aerial videos is available on https://surbhi-31.github.io/Aeriform-in-action/.

Recognizing human activities using light-weight and effective machine learning methodologies

Background Human activity recognition is a dynamic and challenging task. It is a large field of research and development. It involves predicting the movement of a person from the raw sensor data using a machine learning model. To accurately detect human activities for e-health systems, several research attempts have been carried out using data mining and machine learning techniques, but there still is room to improve the performance. To this aim, human activities such as walking, standing, laying, sitting, walking upstairs, walking downstairs are predicted using prominent machine learning models. The aim of human activity recognition is examining actions from photos or video clips. This serves as the driving force behind human activity identification systems' aim to accurately classify input data into the relevant activity category. Methods Six machine learning techniques, including decision tree, random forest, linear regression, Naïve bayes, k-nearest neighbour, and neural networks algorithms, were used for human activity recognition. Results The performance of decision tree, random forest, linear regression, Naïve bayes, k-nearest neighbor, and neural network algorithms was assessed with a human activity recognition dataset. From the results, the random forest classifier and neural network gave good results, whereas the Naïve bayes result was not satisfying. Conclusions We classified the SITTING, STANDING, LAYING, WALKING, WALKING_DOWNSTAIRS, WALKING_UPSTAIRS activities with machine learning techniques with 98% of accuracy

Anomaly detection in internet of medical things with artificial intillegence

Internet of things (IoT) becomes the most popular term in the recent advances in Healthcare devices. The healthcare data in the IoT process and structure is very sensitive and critical in terms of healthy and technical considerations. Outlier detection approaches are considered as principal tool or stage of any IoT system and are mainly categorized in statistical and probabilistic, clustering and classification-based outlier detection. Recently, fuzzy logic (FL) system is used in ensemble and cascade systems with other ML-based tools to enhance outlier detection performance but its limitation involves the false detection of outliers. In this paper, we propose a fuzzy logic system that uses the anomaly score of each point using local outlier factor (LOF), connectivity-based outlier factor (COF) and generalized LOF to eliminate the confusion in classifying points as outliers or inliers. Regarding human activity recognition (HAR) dataset, the FL achieved a value of 98.2 %. Compared to the performance of LOF, COF, and GLOF individually, the accuracy increased slightly, but the increase in precision and recall indicates an increase in correctly classified data and that neither true nor abnormal data is classified wrongly. The results show the increase in precision and recall which indicates an increase in correctly classified data. Thus, it can be confirmed that fuzzy logic with input of scores achieved the desired goal in terms of mitigating cases of false detection of anomalous data. By comparing the proposed ensemble of fuzzy logic and different types of local density scores in this study, the outcomes of fuzzy logic presents a new way of elaborating or fusing the different tools of the same purpose to enhance detection performance

Context-aware mutual learning for semi-supervised human activity recognition using wearable sensors

With the increasing popularity of wearable sensors, deep-learning-based human activity recognition (HAR) has attracted great interest from both academic and industrial fields in recent years. Nevertheless, the performance of deep HAR methods highly depends on the quality and quantity of annotations that are not so prone to obtain in HAR. This practical concern raises broad research of semi-supervised HAR. Despite the brilliant achievements, there remain three important issues to be settled: aggravation of overfitting, deviation of distribution and ignorance of contextual information. This paper proposes a novel context-aware mutual learning method for semi-supervised HAR. Firstly, a semi-supervised mutual learning framework is introduced to alleviate the overfitting of single network. In this framework, the main and auxiliary networks are collaboratively trained with supervised information from each other. Secondly, the distribution-preserving loss, which minimizes the distance between the class distribution of predictions and that of labeled data, is proposed to hinder the deviation of the distribution. Finally, the contextual information from the neighbor sequences is adopted through a context-aware aggregation module. This module extracts richer information from a broader range of sequences. Our method is validated on four characteristic published human activity recognition datasets: UCI, WISDM, PAMAP2 and mHealth. The experimental result shows that the proposed method achieves superior performance compared with four typical methods in semi-supervised HAR.

Transfer Learning Approach for Human Activity Recognition Based on Continuous Wavelet Transform

Over the last few years, human activity recognition (HAR) has drawn increasing interest from the scientific community. This attention is mainly attributable to the proliferation of wearable sensors and the expanding role of HAR in such fields as healthcare, sports, and human activity monitoring. Convolutional neural networks (CNN) are becoming a popular approach for addressing HAR problems. However, this method requires extensive training datasets to perform adequately on new data. This paper proposes a novel deep learning model pre-trained on scalograms generated using the continuous wavelet transform (CWT). Nine popular CNN architectures and different CWT configurations were considered to select the best performing combination, resulting in the training and evaluation of more than 300 deep learning models. On the source KU-HAR dataset, the selected model achieved classification accuracy and an F1 score of 97.48% and 97.52%, respectively, which outperformed contemporary state-of-the-art works where this dataset was employed. On the target UCI-HAPT dataset, the proposed model resulted in a maximum accuracy and F1-score increase of 0.21% and 0.33%, respectively, on the whole UCI-HAPT dataset and of 2.82% and 2.89%, respectively, on the UCI-HAPT subset. It was concluded that the usage of the proposed model, particularly with frozen layers, results in improved performance, faster training, and smoother gradient descent on small HAR datasets. However, the use of the pre-trained model on sufficiently large datasets may lead to negative transfer and accuracy degradation.

Multiscale Deep Feature Learning for Human Activity Recognition Using Wearable Sensors

Deep convolutional neural networks (CNNs) achieve state-of-the-art performance in wearable human activity recognition (HAR), which has become a new research trend in ubiquitous computing scenario. Increasing network depth or width can further improve accuracy. However, in order to obtain the optimal HAR performance on mobile platform, it has to consider a reasonable tradeoff between recognition accuracy and resource consumption. Improving the performance of CNNs without increasing memory and computational burden is more beneficial for HAR. In this article, we first propose a new CNN that uses hierarchical-split (HS) idea for a large variety of HAR tasks, which is able to enhance multiscale feature representation ability via capturing a wider range of receptive fields of human activities within one feature layer. Experiments conducted on benchmarks demonstrate that the proposed HS module is an impressive alternative to baseline models with similar model complexity, and can achieve higher recognition performance (e.g., 97.28%, 93.75%, 99.02%, and 79.02% classification accuracies) on UCI-HAR, PAMAP2, WISDM, and UNIMIB-SHAR. Extensive ablation studies are performed to evaluate the effect of the variations of receptive fields on classification performance. Finally, we demonstrate that multiscale receptive fields can help to learn more discriminative features (achieving 94.10% SOTA accuracy) in weakly labeled HAR dataset.

Multi-level feature fusion for multimodal human activity recognition in Internet of Healthcare Things

Human Activity Recognition (HAR) has become a crucial element for smart healthcare applications due to the fast adoption of wearable sensors and mobile technologies. Most of the existing human activity recognition frameworks deal with a single modality of data that degrades the reliability and recognition accuracy of the system for heterogeneous data sources. In this article, we propose a multi-level feature fusion technique for multimodal human activity recognition using multi-head Convolutional Neural Network (CNN) with Convolution Block Attention Module (CBAM) to process the visual data and Convolutional Long Short Term Memory (ConvLSTM) for dealing with the time-sensitive multi-source sensor information. The architecture is developed to be able to analyze and retrieve channel and spatial dimension features through the use of three branches of CNN along with CBAM for visual information. The ConvLSTM network is designed to capture temporal features from the multiple sensors’ time-series data for efficient activity recognition. An open-access multimodal HAR dataset named UP-Fall detection dataset is utilized in experiments and evaluations to measure the performance of the developed fusion architecture. Finally, we deployed an Internet of Things (IoT) system to test the proposed fusion network in real-world smart healthcare application scenarios. The findings from the experimental results reveal that the developed multimodal HAR framework surpasses the existing state-of-the-art methods in terms of multiple performance metrics.

A hybrid tuple selection pipeline for smartphone based Human Activity Recognition

With the increasing use of sensors in almost every device and application worldwide, the volume of data increases exponentially. Instance selection is an important data preprocessing step that can be applied to many machine learning models. It is even crucial for executing machine learning tasks in constrained environments as demanded by many Internet of Things (IoT) enabled applications today such as the Human Activity Recognition (HAR) domain of applications. In this paper, we have proposed a hybrid selection and training pipeline methodology that combines the nearest neighbor concept with evolutionary computing to address the instance selection problem of smartphone sensing-based HAR. A clustering algorithm has been proposed first followed by a Genetic Algorithm based instance selection approach. It is so far the first of its kind in the aforementioned domain of HAR based on smartphone sensing data. The proposed instance selection has been tested on two of the popular benchmark HAR datasets — UCI-HAR and WISDM. The experimented outputs report that the proposed method has effectively reduced the dataset size by around 40% for the benchmark datasets while retaining the recognition accuracy above 94%. This is a clear depiction of the removal of the outliers from the instance set. The proposed approach has also been compared with other state-of-the-art under-sampling approaches and the results show that our approach has performed better.

CapMatch: Semi-Supervised Contrastive Transformer Capsule With Feature-Based Knowledge Distillation for Human Activity Recognition.

This article proposes a semi-supervised contrastive capsule transformer method with feature-based knowledge distillation (KD) that simplifies the existing semisupervised learning (SSL) techniques for wearable human activity recognition (HAR), called CapMatch. CapMatch gracefully hybridizes supervised learning and unsupervised learning to extract rich representations from input data. In unsupervised learning, CapMatch leverages the pseudolabeling, contrastive learning (CL), and feature-based KD techniques to construct similarity learning on lower and higher level semantic information extracted from two augmentation versions of the data", weak" and "timecut", to recognize the relationships among the obtained features of classes in the unlabeled data. CapMatch combines the outputs of the weak-and timecut-augmented models to form pseudolabeling and thus CL. Meanwhile, CapMatch uses the feature-based KD to transfer knowledge from the intermediate layers of the weak-augmented model to those of the timecut-augmented model. To effectively capture both local and global patterns of HAR data, we design a capsule transformer network consisting of four capsule-based transformer blocks and one routing layer. Experimental results show that compared with a number of state-of-the-art semi-supervised and supervised algorithms, the proposed CapMatch achieves decent performance on three commonly used HAR datasets, namely, HAPT, WISDM, and UCI_HAR. With only 10% of data labeled, CapMatch achieves F1 values of higher than 85.00% on these datasets, outperforming 14 semi-supervised algorithms. When the proportion of labeled data reaches 30%, CapMatch obtains F1 values of no lower than 88.00% on the datasets above, which is better than several classical supervised algorithms, e.g., decision tree and k -nearest neighbor (KNN).

MhaGNN: A Novel Framework for Wearable Sensor-Based Human Activity Recognition Combining Multi-Head Attention and Graph Neural Networks

Obtaining robust feature representations from multi-position wearable sensory data is challenging in human activity recognition (HAR) since data from different positions can have unordered implicit correlations. Graph neural networks (GNNs) represent data as structured graphs by mining complex relationships and interdependency via message passing between the nodes of graphs. This paper proposes a novel framework (MhaGNN) that combines GNNs and the multi-head attention mechanism, aiming to learn more informative representations for multi-position HAR tasks. The MhaGNN framework takes the sensor channels from multiple wearing positions as nodes to construct graph-structured data from the spatial dimension. Besides, the multi-head attention mechanism is introduced to complete the message passing and aggregation of the graphs for spatial-temporal feature extraction. The MhaGNN learns correlations among sensor channels that can be used as compensatory features together with the captured features from each single sensor channel to enhance HAR. Experimental evaluations on three publicly available HAR datasets and a ground-truth dataset demonstrate that our proposed MhaGNN achieves state-of-the-art recognition performance with the captured rich features, including PAMAP2, OPPORTUNITY, MHAEATH and MPWHAR.

DMSTL: A Deep Multi-Scale Transfer Learning Framework for Unsupervised Cross-Position Human Activity Recognition

Human activity recognition (HAR) based on wearable sensors has been a prosperous research topic in recent years. Considering that the sensor may be worn at diverse body positions, obtaining enough labeled human activity data for each body-worn position is usually expensive and labor-intensive. Furthermore, the variability and diversity of data distribution induced by different body-worn positions make the HAR model trained on data collected from one body position perform poorly for other body-worn positions. To achieve accurate HAR with low labeling cost, which we call unsupervised cross-position HAR, in this article, we propose a deep multiscale transfer learning (DMSTL) model. In the model, we first introduce an unsupervised source selection method to select the most similar source domain for transferring domain knowledge. Then, we develop a multiscale spatial–temporal Net (MSSTNet) to learn comprehensive multimodal representations from multiple feature subspaces. Finally, we design the category-level adaptation module and the domain-level adversarial module for learning domain-invariant features. We conduct extensive experiments on three public HAR data sets and demonstrate the reasonable generalization performance of the DMSTL, which remarkably outperforms other state-of-the-art baselines.