Benchmark analysis of deep learning algorithms for edge-based robotic applications

Driven by the increasing demand for deploying Artificial Intelligence (AI) in resource-constrained computing devices, this study presents a comprehensive benchmark comparing the inference speed, energy efficiency, and hardware utilization of various deep learning algorithms across multiple commercial edge devices, consisting of Jetson Nano, Jetson Tegra X2 (TX2), Raspberry Pi 4 (RPi4), and Coral Artificial Intelligence Accelerator (Coral AI). Specifically, vision algorithms crucial for robotics applications including object detection, image segmentation, human pose estimation, and face detection are investigated. We provide an analysis of hardware performance across various machine learning frameworks, such as Tensor RunTime (TensorRT), TensorFlow Lite, PyTorch, and Open Source Computer Vision Library (OpenCV). The result highlights the vital role of Graphic Processing Units in enhancing inference speed, yet Central Processing Units remain essential for effective data processing tasks. Furthermore, our study shows that inference speed can be significantly improved by utilizing model optimization techniques and AI accelerator.