Deep Learning-Based Object Detection, Localisation and Tracking for Smart Wheelchair Healthcare Mobility

Area of Science:

• Robotics
• Computer Vision
• Human-Computer Interaction

Background:

• Smart electric wheelchairs aim to increase independence for disabled individuals.
• Current systems often lack sophisticated environmental perception capabilities.
• Assisted navigation requires accurate detection and tracking of key environmental features.

Purpose of the Study:

• To develop an Advanced Driver Assistance System (ADAS) for smart electric wheelchairs.
• To enhance the autonomy of disabled people through improved wheelchair navigation.
• To create a perception layer for detecting and tracking indoor objects like doors and door handles.

Main Methods:

• Adaptation of the YOLOv3 object detection algorithm for identifying doors and handles.
• Depth estimation using an Intel RealSense camera.
• 3D object tracking implemented with the SORT algorithm.

Main Results:

• Successful detection, depth estimation, and 3D tracking of doors and door handles in an indoor environment.
• Development of a short-lifespan semantic map for immediate surroundings.
• Validation through experiments using a custom dataset in a controlled setting.

Conclusions:

• The developed ADAS provides a crucial perception layer for smart electric wheelchairs.
• This system significantly contributes to improving the autonomy and navigation capabilities of disabled users.
• The integration of object detection, depth estimation, and tracking offers a robust solution for indoor wheelchair assistance.