Something Never Before Seen.
One of the core facts we understood from the very beginning is that a smartphone and most motion tracking devices are not build to capture highly accurate data. While those devices are equipped with sensors, e.g. location, accelerometer, magnetometer, gyroscope or other sensors, the capability of producing a high quality, homogenous data stream is very limited. Modern smartphones, for example, are primarily utilizing motion sensors to either rotate pictures as the device is rotated or to detect repetitive motion sequences such as walking or running.
At Tourmaline Labs™, we studied the challenge and decided to solve it. The challenge is to utilize a substandard device, the smartphone, with low quality sensors that are poorly calibrated and drifting over time to derive a high quality, high frequency, and homogenous mobility data stream without changing the user experience in terms of battery drain or requiring the device to be mounted inside a cradle.
The key to success was to fundamentally understand the challenge in front of us and to assemble a team of world-class talent that is uniquely qualified to address the problem at hand. The initial team spans from aerospace engineers with a background in embedded system intelligence and small satellite tracking to computer scientists with deep knowledge in artificial intelligence and machine learning to embedded software engineers with an appreciation for advanced algorithmic implementations inside a resource constraint environment. This team then combined techniques ranging from analog signal processing, convex optimization, bio-metrics, system identification and machine learning to turn a noisy, non calibrated, and not oriented sensor data stream into a high quality, fully normalized, steady, and homogenous mobility data stream.
There are a couple of key secrets to this success. First, we decomposed the problem without breaking any interconnected dependencies that would alter the nature of the problem and preserve the physical and logical constrains. In a second step, we applied techniques from different disciplines to the individual sub-problems that were most suited while ensuring the assumptions each technique would build on are not violated by any other technique used. Last but not least, we designed the system to be self-calibrating and self-learning – an artificial intelligence techniques that enable to system to evolve and adapt to variations of the same underlying problem without any human assistance. This enables the system to reason over time; to learn more about each individual device or data stream; to transfer knowledge between different points in time or space; and evolve as mobility is evolving.
As a result, our patented drive behavior analytics platform, aggregates and processes data from over 4200+ different device types in over 150+ countries and achieve a level of data quality and integrity that is otherwise only possible with high-cost, manually calibrated, and dedicated sensor equipment deployed in a controlled environment.