Preliminary Analysis of Collar Sensors for Guide Dog Training Using Convolutional Long Short-Term Memory, Kernel...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Animal Behavior

Background:

• Guide dogs significantly enhance independence for visually impaired individuals.
• High demand and training costs limit the availability of guide dogs.
• Understanding guide dog challenges is crucial for optimizing their training and performance.

Purpose of the Study:

• To develop and evaluate a multi-sensor smart collar system for analyzing guide dog behavior and sensor data.
• To compare the efficacy of different machine learning models for processing sensor data.
• To identify key data patterns and simplify guide dog testing procedures.

Main Methods:

• Development of a multi-sensor smart collar system for data acquisition.
• Comparison of Convolutional Long Short-Term Memory (Conv-LSTM) and Kernel Principal Component Analysis (KPCA) for supervised learning.
• Utilizing an unsupervised autoencoder to create a lexicon of data patterns.
• Analysis of data from inertial measurement units and environmental acoustic sensors.

Main Results:

• Conv-LSTM and KPCA achieved approximately 40% accuracy on a 10-state system using optimized data.
• Inertial measurement units provided the most significant predictive information.
• Environmental acoustic sensing data offered minor performance improvements.
• An unsupervised autoencoder identified distinct data patterns and potential for state simplification.

Conclusions:

• The smart collar system provides valuable data for understanding guide dog behavior and optimizing training.
• Machine learning models, particularly Conv-LSTM and KPCA, can effectively analyze sensor data.
• Simplifying testing states by combining them into superstates is feasible.
• Future research can leverage this system to further enhance guide dog training and support.