Neural network-based algorithm for door handle recognition using RGBD cameras.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Robot adaptability requires effective interaction with diverse objects like doorknobs.
• Current robotic systems face challenges in recognizing and manipulating varied doorknob designs in different environments.

Purpose of the Study:

• To develop and implement a robust method for robots to recognize door handle positions using RGBD camera data.
• To enhance autonomous robot capabilities for real-world interaction with common household objects.

Main Methods:

• A dataset of 5000 annotated doorknob images was created.
• A neural network architecture based on MobileNetV2 with a custom decoder and a novel activation function was employed.
• Real-time image processing was a key implementation goal.

Main Results:

• The proposed approach successfully identifies and localizes door handles from RGBD camera input.
• The model achieves real-time processing speeds of up to 16 images per second.
• The custom architecture and activation function contribute to increased accuracy and efficiency.

Conclusions:

• The developed method significantly advances robot perception for doorknob manipulation.
• This research contributes to the practical deployment of autonomous robots in diverse settings.
• The real-time performance enables more fluid and effective human-robot interaction.