Deep Learning based Driving Posture Stability Analysis for People with Mobility Challenges

Mobility scooters are critical in facilitating social participation of people with mobility challenges and thus improving their life quality. However, the safety issues of driving mobility scooter are real concerns and may not be assessed in a timely fashion for many patients. In this paper we enable driving posture stability analysis at home or in the community setting for people with mobility challenges by using only video recordings of their driving. In particular, we design a system that extracts upper body keypoints’ 2D coordinates from video frames and builds an autoencoder model to perform stability analysis. We explore two architectures of the autoencoder, with Long Short Term Memory (LSTM) and Convolutional Neural Networks (CNN), respectively emphasizing the temporal and spatial relationships of upper body keypoints in driving movements. Evaluations using patients driving posture data collected have shown that both architectures have achieved over 0.99 Precision-Recall Area Under the Curve (AUC), and 0.8 ROC AUC, indicating excellent model accuracy levels.