Object Detection on TPU Accelerated Embedded Devices

Modern edge devices are capable of onboard processing of computational heavy tasks, such as artificial intelligence-driven computer vision. An increasing number of deep learning-based object detection networks are frequently proposed with lightweight structures to be deployed on mobile platforms without the need for cloud computing.

Comparing these networks is challenging due to the variety in hardware and frameworks and because of different model complexity. This paper investigates models that can be deployed on cross-functional single-board computers without utilizing the power of GPUs. This paves the way towards performing accurate, cheap, and fast object detection, even suited for industrial applications within Industry 4.0.

Four state-of-the-art neural networks are trained via transfer learning, then deployed and tested on the Raspberry Pi 4B and the Coral Edge TPU accelerator from Google as a co-processor. The comparison of the models focuses on the inference time, the versatility of the deployment, training, and finally the accuracy of the retrained networks on a selection of datasets with different feature characteristics.

Our code can be found in the following repository: https://github.com/kberci/Deep-Learning-based-Object-Detection.