Tracking by segmentation with future motion estimation applied to person-following robots.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Person-following is essential for service robots, with vision technology being key for environmental understanding.
• Existing tracking-by-detection methods require large datasets and are vulnerable to environmental noise.

Purpose of the Study:

• To develop a novel real-time tracking-by-segmentation approach for robust person-following in service robots.
• To enhance tracking accuracy by utilizing pixel-level segmentation and future motion estimation.

Main Methods:

• A single-shot segmentation tracking neural network for precise foreground segmentation, overcoming region of interest (ROI) limitations.
• A classification-lock pre-trained layer to constrain feature outliers and a discriminative correlation filter to prevent misrecognition.
• A motion estimation neural network to predict the target's future motion for robot control.

Main Results:

• The proposed framework demonstrates effectiveness across VOT, LaSot, YouTube-VOS, and Davis tracking datasets.
• Achieved precise pixel-level tracking and accurate future motion prediction for the target individual.
• Successfully validated for long-term person-following tasks in indoor environments.

Conclusions:

• The real-time tracking-by-segmentation framework offers a significant improvement over traditional methods for robot person-following.
• This approach shows strong potential for practical application in service robots, enhancing their ability to interact with humans.