Recent developments to Tensorflow on edge devices has allowed to squeeze in a Tensorflow Processing Unit into one small package. xFarm Tech has used Google Coral TPU to accelerate inference on a Raspberry Pi 4. A total of 25 frames per second was achieved on two camera feeds using one Coral TPU. That is very impressive indeed and serves as a successful proof-of-concept for inference on the edge system-on-chip such as a Raspberry Pi. The model trained was based on Mobilenet V2 converted to a tflite to perform object detection of person and person with personal protective equipment such as a helmet.
xFarm Tech has also tested PoseNet, pose estimation using tensorflow lite, on the Google Coral TPU.