Abstract:Robots have become increasingly prevalent in various aspects of daily life. In an increasing number of scenarios, particularly in social contexts, such as those involving delivery robots and self-driving buses, human-robot interaction (HRI) is essential for conveying decision commands to the robot's control algorithm. The growing convergence of human activity spaces and robot workspaces presents increasingly emergent and challenging interaction scenarios. This transformation necessitates convenient measurement sensor and gesture recognition algorithms with shorter recognition delays and more reliable recognition results. In pursuit of this objective, we propose a novel pipeline for continuous and rapid gesture recognition employing a portable inertial measurement unit (IMU) device. First, we design a lightweight neural network (NN) to distinguish the gesture from daily hand movement (gesture spotting). Then, one lookback window evaluation is inserted between gesture spotting and recognition, aiming to reduce the false positive rate of gesture spotting. Finally, a novel transformer-based model is used to recognize expected hand gestures (gesture recognition), which can consider both cross-time and cross-sensor dependencies. In this study, the IMU data of six gestures are collected from 20 participants for training and validation. The experimental results show that the proposed gesture spotting model detects all gesture movements in the test data with a 0.3% false positive rate. The gesture recognition model achieves better accuracy than multiple state-of-the-art models. Overall, this study proposes an efficient and reliable pipeline for single IMU-based continuous gesture recognition for practical applications.

Continuous Human Action Detection Based on Wearable Inertial Data

Human Action Recognition Using Deep Learning Methods.

Human Action Recognition From Digital Videos Based on Deep Learning.

Inertial Sensor Data To Image Encoding For Human Action Recognition

Human Action Recognition Based on Hierarchical Multi-Scale Adaptive Conv-Long Short-Term Memory Network

Human Activity Recognition with Convolutional Neural Netowrks

A Novel Deep Learning Bi-GRU-I Model for Real-Time Human Activity Recognition Using Inertial Sensors

WristCam: A Wearable Sensor for Hand Trajectory Gesture Recognition and Intelligent Human–Robot Interaction

Attention-based LSTM Network for Wearable Human Activity Recognition

Introducing IHARDS-CNN: A Cutting-Edge Deep Learning Method for Human Activity Recognition Using Wearable Sensors

Continuous Hand Gestures Detection and Recognition in Emergency Human-Robot Interaction Based on the Inertial Measurement Unit

Continuous Human Action Recognition in Real Time

Human Activity Recognition Based on Wireless Electrocardiogram and Inertial Sensors

Hierarchical Multi-Scale Adaptive Conv-LSTM Network for Human Action Recognition Based on Wearable Sensors.

Decoding Human Activities: Analyzing Wearable Accelerometer and Gyroscope Data for Activity Recognition

Deep Learning-Based Human Action Recognition in Videos

Human Activity Recognition using Deep Learning Models on Smartphones and Smartwatches Sensor Data

Comparison Study of Inertial Sensor Signal Combination for Human Activity Recognition based on Convolutional Neural Networks

Iss2Image: A Novel Signal-Encoding Technique for CNN-Based Human Activity Recognition

Real-Time Onboard Human Motion Recognition Based On Inertial Measurement Units

Continuous Fine-Grained Arm Action Recognition Using Motion Spectrum Mixture Models