When building a unified vision system or gradually adding new capabilities to a system, the usual assumption is that training data for all tasks is always available. However, as the number of tasks grows, storing and retraining on such data becomes infeasible. A new problem arises where we add new capabilities to a Convolutional Neural Network (CNN), but the training data for its existing capabilities are unavailable. We propose our Learning without Forgetting method, which uses only new task data to train the network while preserving the original capabilities. Our method performs favorably compared to commonly used feature extraction and fine-tuning adaption techniques and performs similarly to multitask learning that uses original task data we assume unavailable. A more surprising observation is that Learning without Forgetting may be able to replace fine-tuning with similar old and new task datasets for improved new task performance.

Learning without Forgetting

The vast proliferation of sensor devices and Internet of Things enables the applications of sensor-based activity recognition. However, there exist substantial challenges that could influence the performance of the recognition system in practical scenarios. Recently, as deep learning has demonstrated its effectiveness in many areas, plenty of deep methods have been investigated to address the challenges in activity recognition. In this study, we present a survey of the state-of-the-art deep learning methods for sensor-based human activity recognition. We first introduce the multi-modality of the sensory data and provide information for public datasets that can be used for evaluation in different challenge tasks. We then propose a new taxonomy to structure the deep methods by challenges. Challenges and challenge-related deep methods are summarized and analyzed to form an overview of the current research progress. At the end of this work, we discuss the open issues and provide some insights for future directions.

Deep Learning for Sensor-based Human Activity Recognition: Overview, Challenges and Opportunities

A review and categorization of techniques on device-free human activity recognition

Recognition of human movements with radar for ambient activity monitoring is a developed area of research that yet presents outstanding challenges to address In real environments, activities and movements are performed with seamless motion, with continuous transitions between activities of different duration and a large range of dynamic motions, compared with discrete activities of fixed-time lengths which are typically analysed in the literature This paper proposes a novel approach based on recurrent LSTM and Bi-LSTM network architectures for continuous activity monitoring and classification This approach uses radar data in the form of a continuous temporal sequence of micro-Doppler or range-time information, differently from from other conventional approaches based on convolutional networks that interpret the radar data as images Experimental radar data involving 15 participants and different sequences of 6 actions are used to validate the proposed approach It is demonstrated that using the Doppler-domain data together with the Bi-LSTM network and an optimal learning rate can achieve over 90% mean accuracy, whereas range-domain data only achieved approximately 76% The details of the network architectures, insights in their behaviour as a function of key hyper-parameters such as the learning rate, and a discussion on their performance across are provided in the paper

/pdf/continuous-human-activity-classification-from-fmcw-radar-4a15x2ocmk.pdf

Continuous Human Activity Classification From FMCW Radar With Bi-LSTM Networks

Human Action Recognition (HAR), aiming to understand human behaviors and then assign category labels, has a wide range of applications, and thus has been attracting increasing attention in the field of computer vision. Generally, human actions can be represented using various data modalities, such as RGB, skeleton, depth, infrared sequence, point cloud, event stream, audio, acceleration, radar, and WiFi, etc., which encode different sources of useful yet distinct information and have various advantages and application scenarios. Consequently, lots of existing works have attempted to investigate different types of approaches for HAR using various modalities. In this paper, we give a comprehensive survey for HAR from the perspective of the input data modalities. Specifically, we review both the hand-crafted feature-based and deep learning-based methods for single data modalities, and also review the methods based on multiple modalities, including the fusion-based frameworks and the co-learning-based approaches. The current benchmark datasets for HAR are also introduced. Finally, we discuss some potentially important research directions in this area.

/pdf/human-action-recognition-from-various-data-modalities-a-3k5a13wj1t.pdf

Human Action Recognition from Various Data Modalities: A Review

Radar, as one of the sensors for human activity recognition (HAR), has unique characteristics such as privacy protection and contactless sensing. Radar-based HAR has been applied in many fields such as human–computer interaction, smart surveillance and health assessment. Conventional machine learning approaches rely on heuristic hand-crafted feature extraction, and their generalization capability is limited. Additionally, extracting features manually is time–consuming and inefficient. Deep learning acts as a hierarchical approach to learn high-level features automatically and has achieved superior performance for HAR. This paper surveys deep learning based HAR in radar from three aspects: deep learning techniques, radar systems, and deep learning for radar-based HAR. Especially, we elaborate deep learning approaches designed for activity recognition in radar according to the dimension of radar returns (i.e., 1D, 2D and 3D echoes). Due to the difference of echo forms, corresponding deep learning approaches are different to fully exploit motion information. Experimental results have demonstrated the feasibility of applying deep learning for radar-based HAR in 1D, 2D and 3D echoes. Finally, we address some current research considerations and future opportunities.

https://www.mdpi.com/2072-4292/11/9/1068/pdf

A Survey of Deep Learning-Based Human Activity Recognition in Radar

Deep cascading network architecture for robust automatic modulation classification

Human activity recognition (HAR) plays a vital role in many applications, such as surveillance, in-home monitoring, and health care. Portable radar sensor has been increasingly used in HAR systems in combination with deep learning (DL). However, it is both difficult and time-consuming to obtain a large-scale radar dataset with reliable labels. Insufficient labeled data often limit the generalization of DL models. As a result, the performance of DL models will drop when being applied to a new scenario. In this sense, only labeling a small portion of data in the large-scale radar dataset is more feasible. In this article, we propose a semisupervised transfer learning (TL) algorithm, ``joint domain and semantic transfer learning (JDS-TL),'' for radar-based HAR, which is composed of two modules: unsupervised domain adaptation (DA) and supervised semantic transfer. By employing a sparsely labeled dataset to train the HAR model, the proposed method alleviates the need of labeling a significantly large number of radar signals. We adopt a public radar micro-Doppler spectrogram dataset including six human activities to evaluate JDS-TL. Experiments show that the proposed JDS-TL is able to recognize the six activities with an average accuracy of 87.6% when there are only 10% instances labeled in the training dataset. Ablation analysis also demonstrates the efficiency of the DA and the semantic transfer modules.

Semisupervised Human Activity Recognition With Radar Micro-Doppler Signatures

The performance of deep learning (DL) algorithms for radar-based human motion recognition (HMR) is hindered by the diversity and volume of the available training data. In this article, to tackle the issue of insufficient training data for HMR, we propose an instance-based transfer learning (ITL) method with limited radar micro-Doppler (MD) signatures, alleviating the burden of collecting and annotating a large number of radar samples. ITL is a unique algorithm that consists of three interconnected parts, including DL model pretraining, correlated source data selection, and adaptive collaborative fine-tuning (FT). Any of the three components cannot be excluded; otherwise, the performance of the entire algorithm decreases. The experiments with a radar data set of six human motions show that ITL achieves state-of-the-art performance for HMR with limited training samples, outperforming several existing transfer learning approaches. Especially, when there are only 100 samples per person per class, ITL yields an F1 score of 96.7%. Last but not least, ITL is more generalized to human motion differences. Though adapted to recognize the persons’ motions in a small-scale target data set, ITL can also classify the persons’ motion data used for pretraining, achieving up to 11.0% F1 score enhancement over the conventional FT method.

/pdf/human-motion-recognition-with-limited-radar-micro-doppler-njv1nz7t54.pdf

Human Motion Recognition With Limited Radar Micro-Doppler Signatures

Short-range radar has become one of the latest sensor technologies for the Internet of Things (IoT), and it plays an increasingly vital role in IoT applications. As the essential task for various smart-sensing applications, radar-based human activity recognition and person identification have received more attention due to radar’s robustness to the environment and low power consumption. Activity recognition and person identification are generally treated as separate problems. However, designing different networks for these two tasks brings a high computational complexity and wastes of resources to some extent. Furthermore, there are some correlations in activity recognition and person identification tasks. In this work, we propose a multiscale residual attention network (MRA-Net) for joint activity recognition and person identification with radar micro-Doppler signatures. A fine-grained loss weight learning (FLWL) mechanism is presented for elaborating a multitask loss to optimize MRA-Net. In addition, we construct a new radar micro-Doppler dataset with dual labels of activity and identity. With the proposed model trained on this dataset, we demonstrate that our method achieves the state-of-the-art performance in both radar-based activity recognition and person identification tasks. The impact of the FLWL mechanism was further investigated, and ablation studies of the efficacy of each component in MRA-Net were also conducted.

Xinyu Li

Papers

A Survey of Deep Learning-Based Human Activity Recognition in Radar

Deep cascading network architecture for robust automatic modulation classification

Semisupervised Human Activity Recognition With Radar Micro-Doppler Signatures

Human Motion Recognition With Limited Radar Micro-Doppler Signatures

A Mutiscale Residual Attention Network for Multitask Learning of Human Activity Using Radar Micro-Doppler Signatures