University of Twente Student Theses


Evaluating the performance of simulated IMU data for animal activity recognition

Vanwinsen, W. (2021) Evaluating the performance of simulated IMU data for animal activity recognition.

[img] PDF
Abstract:Labeled data for IMU-based animal activity recognition (AAR) is difficult to collect due to major time- and resource constraints. As a result, the datasets which can be used for machine learning are often small and mostly cover domesticated animals such as cattle and horses. This scarcity of data hinders the development of robust predictive models for animal activity. To overcome this issue, researchers have proposed methods to generate virtual IMU-sensor streams from videos by using pose estimation and forward kinematics. The underlying idea being that this simulated IMU data is easier to obtain and label. This paper describes a theoretical framework for simulating IMU data from videos of animals. This paper determines the viability of using simulated IMU data for activity recognition. Acceleration and angular rate estimates are derived from optical motion capture data of horses and compared to real IMU data through signal-level evaluation. The real and simulated data is pre-processed to increase their correlation. Additionally, several machine learning models (SVM, Random Forest, KNN and Naive Bayes) are trained using real, simulated and a combination of both to assess the effects on model performance. In the latter case, augmentation of real data with simulated data can increase the average F1-score by more than 20\% when little real data is available for training. Classifiers trained using only simulated data show predictive performance that is inferior to that of classifiers trained using real data. Nevertheless, competitive recognition accuracy can be achieved.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Creative Technology BSc (50447)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page