Identifying and tracking a target in an indoor environment : algorithms and embedded systems

Abstract

There are a lot of people with limited lower body function that are dependent on wheelchairs for their mobility. Many of these people can still walk but are discouraged from walking away from their wheelchairs without support. These people could benefit from a self-driving wheelchair that follows them around at a relatively short distance since the individual needs to have the ability to get back in the wheelchair at any instance. The purpose of this research was to develop a cost-effective component of a system that can track and follow a person in real time using computer vision. This component should be able to identify and track the subject outputting directional data to navigate the wheelchair in close proximity to the moving subject. The component utilizes a Jetson Nano as the processing unit and an Arducam camera for video input. To achieve our goal we trained a YOLOv4-tiny object detection model using a combination of existing data and data collected and labeled specifically for this research. The model was capable of detecting and identifying a human target wearing the AR-tag on their back. We show that our component is able to successfully track and follow a person wearing an AR-tag in real time, providing reliable directional output using only moderate computing power. The component was evaluated using Nvidia Jetson Nano where we got at least one prediction per second, at a distance up to 3.5 meters with accuracy being over 98\%. Such hardware is low-power and can reasonably be operated on batteries and thus deployed on a mobile platform such as a wheelchair. From the experiments made, we can conclude that a low-cost human following robot is possible using the YOLOv4-tiny object detection model. Further work is needed to improve the navigation and optimization of the detection model, additionally.