Deep Learning-Based Subject Independent Human Activity Recognition using Smart Lacelock Data

• Published in: 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) Vol. 2024; pp. 1 - 4
• Main Authors: Neya, Najmeh Movahhed, Sazonov, Edward, Shen, Xiangrong
• Format: Conference Proceeding Journal Article
• Language: English
• Published: United States IEEE 01.07.2024
• Subjects: Accuracy; CNN; Convolutional neural networks; Data models; Deep Learning; Footwear; Human Activities; Human activity recognition; Human Activity Recognition (HAR); Humans; IMU; Inertial navigation; Leave-One-Out (L1O) method; Legged locomotion; Measurement units; Medical services; Motion Capture - instrumentation; Pattern Analysis, Machine; Smart Lacelock Device; Stair Climbing; Stairs; Subject Independent; Walking; Wearable Electronic Devices
• ISSN: 2694-0604
• DOI: 10.1109/EMBC53108.2024.10781739

Human Activity Recognition (HAR) field is rapidly growing and the classification of human activities based on sensor data is crucial for applications in healthcare, rehabilitation and numerous other sectors. In this paper we use a novel device and attempt Deep Learning-based HAR from the device data.Typically, sensor-based HAR tasks use data from accelerometer and gyroscope within Inertial Measurement Units (IMU). But in this work we use the data from Smart Lacelock device, which is home to IMU and loadcell, introducing an additional sensor, aimed at complementing IMU data. This novel device ensures user comfort by attaching to the user's shoe as a shoelace tensioning device without any shoe modification. The data for this study was collected by the UA HuB-Robotics Lab from eight participants.Using this comprehensive dataset, we propose a CNN based model to classify activities such as walking, stair climbing, and stair descending. The model comprises three consecutive CNN blocks, and within each block there is a convolutional layer, a max-pooling layer, a Rectified Linear Unit (ReLU) layer, and a normalization layer. The model has a dropout and a flatten layer right after the third block of CNN and concludes with 2 dense layers. Our model achieves an average recognition accuracy of 98.4% using the leave-one-out (L1O) technique.In this work Smart Lacelock device demonstrated feasibility in recognition of a set of human activities and the results support further investigation of its applications in HAR.