Semester of Graduation

Spring 2026

Degree

Master of Science in Computer Science (MSCS)

Department

Computer Science and Engineering

Document Type

Thesis

Abstract

Active exoskeletons are being developed to support human movement in physically demanding industries such as construction. For these systems to work effectively, they must be able to correctly identify the user’s current activity. This process is known as locomotion mode detection and plays an important role in selecting the appropriate control parameters for exoskeletons. Many existing approaches use inertial measurement units (IMUs) to recognize these activities and have shown strong performance. However, most of these methods depend on large amounts of labeled data collected under specific conditions. As a result, they often do not perform well when applied to new users, different environments, or changes in sensor configurations. Collecting and labeling new data for every scenario is time consuming and not practical for real-world deployment.

To address this limitation, this thesis presents a pretraining-based framework for creating efficient and adaptable locomotion mode classifier by enabling rapid on-device training. In this approach, a single IMU-based feature extraction backbone is first trained using large amounts of unlabeled IMU data. This pretraining stage allows the backbone to learn general patterns of human motion without requiring manual labeling. After this step, a lightweight classifier head is added on top of the pretrained feature extraction backbone to form an end-to-end classifier. Only classifier head is trained using a small amount of labeled data collected from the target user, setting or environment that allows the model to be calibrated to the target conditions. The proposed method is designed to run on embedded devices and supports on-device training that makes it suitable for personalized applications.

The performance of the proposed framework is evaluated on commercially available embedded platforms including the NVIDIA Jetson Orin Nano and Raspberry Pi 5. Experimental results show that the proposed approach achieves high classification accuracy reaching up to 98\% and comparable to fully supervised training methods. However, our proposed approach takes significantly less training time with reduction of approximately 15 times on the Jetson platform and 75 times on the Raspberry Pi. Overall, this work is a step towards the direction of off-the-shelf exoskeletons products for industries like constructions.

Date

3-27-2026

Committee Chair

Lao, Dong

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Share

COinS