Data Fusion for Cross-Domain Real-Time Object Detection on the Edge

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Edge computing enables real-time robot control by processing data locally.
• Running multiple neural networks on a single node presents resource allocation challenges.

Purpose of the Study:

• To evaluate the feasibility of a unified neural network model for robot control.
• To compare resource efficiency and accuracy against separate, specialized models.

Main Methods:

• Implemented a single YOLOv5 convolutional neural network for both object detection and hand-gesture classification.
• Compared performance metrics (resource usage, accuracy) against using two independent models.

Main Results:

• Resource utilization was reduced by approximately 50% with the single-model approach.
• Accuracy improved for many classes when using the consolidated model trained on more data.
• Augmenting small datasets with instances from larger ones boosted detection accuracy.

Conclusions:

• A single, multi-task neural network is an efficient solution for edge-computing robot control.
• Consolidating tasks into one model reduces computational load without sacrificing, and often improving, accuracy.
• Dataset augmentation strategies are crucial for optimizing performance with limited data.