Optimization of Trash Identification on the House Compound Using a Convolutional Neural Network (CNN) and Sensor...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Current object identification methods struggle to differentiate between 3D objects and 2D images.
• This limitation leads to errors, such as robotic grippers attempting to grasp non-existent objects (e.g., pictures of trash).
• Accurate 3D object detection is crucial for effective waste management and robotic manipulation.

Purpose of the Study:

• To optimize object identification, specifically for distinguishing and handling trash in domestic environments.
• To develop a method that can reliably differentiate between real (3D) and image-based (2D) trash objects.
• To improve the efficiency and accuracy of robotic systems in waste collection tasks.

Main Methods:

• Proposed the Sequential Camera-LiDAR (SCL) method, integrating Convolutional Neural Networks (CNNs) with LiDAR technology.
• Evaluated four CNN architectures (AlexNet, VGG16, GoogleNet, ResNet18) for trash identification accuracy.
• Utilized LiDAR for precise 3D scanning to determine object dimensionality (2D vs. 3D) and inform robotic gripper actions.

Main Results:

• CNN architectures achieved high accuracy in object identification, with GoogleNet reaching 98.3% and ResNet18 reaching 97.5%.
• The SCL method successfully distinguishes between 3D and 2D objects, preventing erroneous grasping of non-real trash.
• Time efficiency improved by 13.33% to 59.26%, with larger objects yielding greater time savings.

Conclusions:

• The combination of CNNs and LiDAR sensors provides accurate trash object identification and 3D/2D discrimination.
• The SCL method enables informed decisions for robotic waste collection, improving efficiency and preventing errors.
• This optimized approach enhances the practical application of robotics in waste management.