Tools for time series data collection

July 2, 2022

While building custom machine learning models for edge devices, we found that high quality training data for real-world events was incredibly difficult to capture. Teaching a system to detect things like glass smashing, forced entry, or the subtle motion patterns a parked vehicle experiences when large trucks or trains pass nearby required bespoke datasets collected directly from physical sensors in uncontrolled environments. Existing tooling was fragmented, unreliable, and difficult for non-engineers to operate in the field.

To solve this, we built Tempo Studio, a suite of data collection and annotation tools designed to support our machine learning workflows. Tempo Studio gave data collection and QA teams the ability to observe live sensor streams, capture synchronized events, annotate samples, and review data quality in real time. The system streamlined the process of gathering and validating large-scale training datasets, helping the team iterate on models significantly faster.

The platform

We built a mobile app for data collection, a dashboard to monitor sensor recordings (image, audio, and motion data) and the low-level software required to pull data off of development kits and boards like Raspberry Pi. They worked together in sync to to help create an end to end annotation and testing suite for real-world AI machine learning development teams.

FieldKits

We built bespoke hardware kits that had the ability to collect high quality audio, motion, and image data in all sorts of real-world environments. We explored how they could be mounted to record important environmental sound and motion signatures and also created a flexible casing to allow for swapping sensor compontents.

In the real-world

Tempo Studio was used to collect automotive security data for edge-ai companies and teams at Ford and Syntiant. Our mobile app was used to control the annotation and recording intervals of the hardware and sensors placed in the vehicle as security events were performed in repetition.