For Google Cloud Next 2020, Google's annual Cloud Platform developer conference, we teamed up to build experiences showcasing live machine learning, scalable infrastructure, and how to leverage the cloud to build distributed systems that bring visibility and predictive information to the fingertips of developers.
Fleet managers and drivers within the fleet face challenges in the information economy. Networked vehicles constantly produce information that overwhelms administrators, and prevents both drivers and managers from receiving actionable insight. Using Google Cloud products -- PubSub, Storage, Firebase, BigQuery, and Tensorflow -- we built a live system for creating data from cars and for visualizing and predicting results.
Using remote controlled cars and raspberry pi augmented with a Google Edge TPU USB accelerator, we created custom machine learning models for recognizing stop signs and speed limits, and built a sensor platform with additional data to create a miniature fleet of vehicles.
RFID cards record the vehicles location throughout the space, determining whether the driver has reached a destination or provide general waypoints. An accelerometer + gyroscope + magnetometer combo is used for "Dead Reckoning" and to perform basic physics calculation for speed, distance, and time. Alongside this navigational data, CPU temperature and vibration replicate real-world performance metrics like engine temp and maintenance needs. From a camera feed, custom Tensorflow models are using to capture lane departure, stop signs, stop lights, pedestrians, and speed limits.
This information is compiled onboard the vehicle and published to PubSub, where it is ingested and sent to Firebase and BigQuery databases for storage and analysis. Here it is used to predict maintenance, safety, and metrics for the fleet. Aggregates are created and sent to a front end for display.
During the COVID outbreak, we quickly pivoted this experience to exist online only. Working with client teams, a system was put in place to simulate a fleet of vehicles. Using Cloud Functions, Dataflow, BigQuery and Firebase databases, we plotted many vehicles onto a map and created a guided walkthrough of capabilities. Some predictive metrics were also included, like brake failure, vehicle health, and driver safety.