Benchmarking Object Detection Deep Learning Models in Embedded Devices.

Area of Science:

• Robotics
• Computer Vision
• Embedded Systems

Background:

• Deep learning (DL) models achieve high accuracy in object detection but demand significant computational resources.
• Robotic platforms often use embedded devices with limited power, necessitating model optimization for deployment.
• Accelerating DL applications on embedded devices requires leveraging specialized AI co-processors.

Purpose of the Study:

• To benchmark the performance of various deep learning object detection models on embedded devices.
• To provide hardware selection guidelines for embedded AI applications.
• To analyze the impact of model quantization and AI co-processor utilization on performance.

Main Methods:

• Benchmarking DL object detection models across two selected embedded boards.
• Evaluating five quantization levels, including CPU-based execution as baseline and AI co-processor acceleration.
• Detailed explanation of the benchmarking procedure and data analysis.

Main Results:

• Quantization and AI co-processor usage significantly improve DL model performance on embedded devices.
• Performance gains vary depending on the model architecture and quantization strategy.
• Hardware selection impacts the feasibility and efficiency of embedded object detection.

Conclusions:

• Optimizing deep learning models through quantization is crucial for embedded robotic applications.
• AI co-processors offer substantial performance benefits over general-purpose CPUs for DL tasks.
• Successful implementation requires careful consideration of hardware capabilities and model constraints.