An Exploration of Machine Learning Methods for Gait Analysis of Potential Guide Dogs

Gait analysis is important for guide dog organizations, as ideal guide dogs have a smooth and efficient gait, where they can also easily shift between and maintain various gaits. Gait quality and natural traveling speed are two of the multiple factors important in matching a guide dog to its visually impaired handler. Gait evaluation typically includes subjective visual observation of the dog or objective assessments obtained from special-designed equipment. Guide dog organizations need a method to easily collect and analyze objective data of gait information. In this work, we explored how various machine learning models could learn and analyze gait patterns from inertial measurements data that were collected during two different data collection experiments using a wearable sensor device. We also evaluated how well each machine learning model could generalize behavior patterns from various dogs under different environments. Additionally, we compared how sensor placement locations could affect gait prediction performance by attaching the sensor device to the dog's neck and back area respectively. The tested machine learning models were able to classify different gaits in the range of 42% to 91% in terms of accuracy, and predict various gait parameters with an error rate ranging from 14% to 29% depending on the setup. Furthermore, we also observed that using behavior data collected from the neck region contains more movement information than the back area. By performing a cross-dataset generalization test on the machine learning models, we found that even with performance drop, the models were able to learn gait-specific behavior patterns that are generalizable for different dogs. Although the results were preliminary, the proposed gait analysis exploration still showed promising potential for studying behavior patterns of candidate guide dogs.