An Efficient Ensemble Binarized Deep Neural Network on Chip with Perception-Control Integrated.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Lightweight Unmanned Aerial Vehicles (UAVs) increasingly utilize deep learning for autonomous navigation.
• Real-time image processing in UAVs demands high computational and storage resources, limiting deployment on edge devices.

Purpose of the Study:

• To develop a computationally efficient deep learning model for UAV navigation.
• To reduce the memory footprint and computational complexity of existing navigation models.

Main Methods:

• Proposed the ensemble binarized DroNet (EBDN) model, integrating binarization and ensemble learning for model compression.
• Developed a novel hardware architecture (EBDNoC) for efficient on-chip implementation of the EBDN model.

Main Results:

• The EBDN model achieved over 7x memory reduction compared to the original DroNet with comparable accuracy.
• The EBDN hardware architecture demonstrated high resource efficiency (10.21 GOP/s/kLUTs) and energy efficiency (208.1 GOP/s/W).

Conclusions:

• The EBDN model offers an effective solution for deploying deep learning-based navigation on resource-constrained UAVs.
• The proposed hardware architecture provides an optimal balance between model performance and hardware resource utilization for embedded systems.