Deep learning model for doors detection a contribution for context awareness recognition of patients with Parkinson’s disease

Abstract

Freezing of gait (FoG) is one of the most disabling motor symptoms in Parkinson’s disease, which is described as a symptom where walking is interrupted by a brief, episodic absence, or marked reduction, of forward progression despite the intention to continue walking. Although FoG causes are multifaceted, they often occur in response of environment triggers, as turnings and passing through narrow spaces such as a doorway. This symptom appears to be overcome using external sensory cues. The recognition of such environments has consequently become a pertinent issue for PD-affected community. This study aimed to implement a real-time DL-based door detection model to be integrated into a wearable biofeedback device for delivering on-demand proprioceptive cues. It was used transfer-learning concepts to train a MobileNet-SSD in TF environment. The model was then integrated in a RPi being converted to a faster and lighter computing power model using TensorFlow Lite settings. Model performance showed a considerable precision of 97,2%, recall of 78,9% and a good F1-score of 0,869. In real-time testing with the wearable device, DL-model showed to be temporally efficient (~2.87 fps) to detect with accuracy doors over real-life scenarios. Future work will include the integration of sensory cues with the developed model in the wearable biofeedback device aiming to validate the final solution with end-users.