Intern - France

Internship – Deep Neural Networks for Classification of Human Motions

 We are seeking a motivated student to join our 2027 Internship Program!

 Overview

At TDK/InvenSense, we’re adding Algorithms for consumer devices (SmartMotion® | TDK InvenSense) using Machine Learning techniques: from sensor data representing some use case, we use Deep Neural Networks as embedded Algorithms that will learn the link from the sensor data to the expected output.

We are testing, in R&D, the approach from real sensor data acquisition to some annotated outputs for several use cases, like Accelerometer based “Bring to See” to wake the watch when the user brings it to his attention, "Wrist Gestures" to navigate in the menus of an e-watch, "Head Gestures" to control smart glasses for incoming calls, "Voice Detection" for smart glasses through Accelerometer based bone conduction …

Some of these new Algorithms (DNN based) are making their path to the market, as part of a Solution offering including the IMU and SW carrying the Algorithm. These successes are quite recent, and we compete with the traditional approaches where a human algorithm developer handcrafts an algorithm. This latter process remains predictable in terms of calendar and performance, usually generates a smaller code, allows to tune the behavior of the feature and sensors, and is best at dynamically managing the sensor system options (such as Low Power)

So, there is still room for improvement and new use cases to use Deep Neural Networks technique, such as how to tune the algorithm, to fit a slightly or significantly different use case than the one in the initial training database, or such as training to get best performances and robustness when the algorithm faces real world, or such as changing sensor settings and ensure good performances, or such as tuning the algorithm for best latency

In new use cases, we don’t have representative data of the classes that we want to classify. Human fall (or Car Crash) are in this case, where we have access to databases representing normal human behavior but have no real Human Fall events, nor Car Crashes events (at least recorded with our sensor).

In the case of Human Fall or Car Crash, we have a device (such as an e-watch or smart band or a mobile phone) that continuously acquires accelerometer and gyroscope data and feeds a Detection/Classification DNN. The DNN (with optional feature extraction) task is to detect the Human Fall or the Car Crash. However, a database of Human Falls or Car Crash is impossible to acquire, so we are willing to test a non-or semi - supervised approach with only everyday life gestures in our Database, learn these normal behaviors and so detect / classify if the new gesture is similar to those in the database.

 Internship Description

In any of the above projects, the internship goal is to bring improvements in using DNN to Detect/Classify Human Motion. We favor a very practical approach, appreciate new visions, methods, and experience.

The Internship would focus on:

• Prepare / Augment a database to generate more diversity before the learning process.

• Find/Design/Train Standard Dense Deep Neural Network Net or Autoencoders. It must be small, accurate, and run in the background on limited memory, and computing power devices. We are using TensorFlow, so scripts and help will be available.

• Using our tools, convert to C Code and embed the algorithm to demonstrate / test the feature

• Improving / adding new features to our DNN library and C converter.

• Implement the feature into a device close to the application (smart watch, mobile phone, …)

•  

Duration: 6 months

Qualifications

• Final year of Engineering school

• Programming and optimization in C/Python, Git.

• Knowledge and interest in machine learning methods.

• Some knowledge or interest in sensors and/or 3D geometry.

• Desire for running experiments with sensor prototypes.

• Pragmatic mindset, autonomy, and creativity.
   

InvenSense, a TDK group company, is a world-leading provider of MEMS and magnetic sensor solutions found in mobile, wearable, smart home, AR/VR, and other consumer and IoT applications. We are innovators with a 20+ year history of patented sensor technologies. As experts in our field, we lead industry breakthroughs in consumer, industrial and automotive technology and reliably deliver high-quality solutions to our partners across global markets.Learn more about us at invensense.tdk.com